04-10 00:01 · 中东,局势,BBC,中东
Both sides have reason to end the war but share no common ground.
Ceasefire or no ceasefire, the Middle East's reshuffling is not yet done
The best hope for the ceasefire talks in Pakistan is that both the United States and Iran have strong reasons to call a halt to the war. The biggest obstacle to their success is a total absence of trust, no discernible common ground and the fact that Israel, America's full partner in the war, has hugely escalated its onslaught on Lebanon.
US President Donald Trump is already speaking about the war in the past tense. He has declared victory and needs an exit. Not only does he have a state visit from King Charles in the diary for later this month, followed by a summit with China's President Xi Jinping in May, there are midterm elections in November. With America's summer holiday season looming, Trump also needs petrol prices to fall back to where they were before he went to war. Royal visits, summits and elections do not mix well with wars.
Iran's regime has its own reasons to end the war. It is as defiant as ever, still able to launch missiles and drones, with its social media warriors pouring out AI videos lampooning Donald Trump. But Iran has suffered massive damage. Cities have come to an economic standstill and the regime needs time to regroup and will try to use the talks in Pakistan to strengthen its position.
The Pakistani intermediaries who will be shuttling between the two delegations have a tough job on their hands. The declared positions of the two sides are as far apart as it is possible to be.
Trump has a 15-point plan that has not been published but leaked versions make it sound more like a surrender document than a basis for negotiation. Iran's 10-point plan contains a list of demands that America has consistently rejected in the past.
Creating a more durable ceasefire will require some kind of agreement to at least keep talking about the two sides' contradictory lists of intractable issues.
It would be hard enough to work through them in peacetime.
In wartime, without any kind of mutual trust, even a form of words that keeps the ceasefire going irrespective of there being no agreement on wider issues will look positive. No agreement at all points towards the road back to war.
The newest, and most urgent problem they face concerns reopening the Strait of Hormuz, the narrow exit from the Gulf. Keeping it closed gives Iran a chokehold on the world economy.
Reopening the waterway that was used by hundreds of ships a day until the US and Israel attacked Iran has become the central issue in the negotiation. The millions of civilians in the Middle East who have been caught up in this conflict hope this negotiation will be the war's endgame.
No victory parade
The Americans did not expect to be spending early April sitting down to ceasefire talks when, alongside Israel on 28 February, they ignited the war with huge strikes that killed among many others Iran's supreme leader, his wife and other members of their family.
Trump was expecting a quick victory, an Iranian version of the US military's stunning kidnap of the Venezuelan leader Nicolas Maduro and his wife in January. Both are on trial in New York on narco-terrorism charges, and the US has installed his former deputy in the presidential palace.
Hopes – expectations – that killing Iran's supreme leader, Ayatollah Ali Khamanei, in the war's first wave of airstrikes would lead to the collapse of the regime were wildly misplaced. His son Mojtaba has not been seen since he was appointed as his successor. There is speculation that he was badly hurt in the attack that killed his parents, as well as reportedly his sister, his wife and one of his sons. With or without the active participation of the new supreme leader, Iran's regime has demonstrated depths of resilience that took Trump by surprise.
Now Donald Trump's representatives, led by his Vice-President JD Vance, must negotiate with adversaries that they claim, incorrectly, to have defeated. "A capital V military victory", as US Defence Secretary Pete Hegseth put it.
The Hormuz factor
The war that the US and Israel ignited is already reshuffling Middle Eastern geopolitics. As the longer-term consequences of the war reveal themselves, that process will deepen. The US and Israel have done immense damage to Iran's armed forces as well as its military and civilian infrastructure. However, while the Iranian regime may be battered, it's also intact. Regime change is not happening. Iran can still launch missiles and drones. That means that despite loud claims, the US and Israel have not translated tactical victories into strategic advances.
Iran, on the other hand, has shown that the closing of the Strait of Hormuz gives it a strategic edge that Donald Trump either dismissed or did not understand when he listened to Israeli Prime Minister Benjamin Netanyahu's arguments for going to war with Iran.
It should not have been a surprise that Iran blocked the Strait when it was attacked. Iran has threatened to do so in the past, and disrupte
▸ 展开全文
04-10 00:01 · 国际,新闻,BBC,国际
Defence Secretary John Healey says there is "no evidence" of any damage to UK infrastructure in the Atlantic.
UK says Russia ran submarine operation over cables and pipelines
Three Russian submarines conducted a "covert" operation over cables and pipelines in waters north of the UK, Defence Secretary John Healey said.
A British warship and aircraft were deployed to deter the "malign" activity by Moscow and there was "no evidence" of any damage to UK infrastructure in the Atlantic, he added.
Addressing Russian President Vladimir Putin directly, Healey said: "We see you. We see your activity over our cables and our pipelines, and you should know that any attempt to damage them will not be tolerated and will have serious consequences."
The Russian embassy in London has denied Healey's claims, according to Moscow's state news agency Tass.
The UK is dependent on its undersea cables and pipelines for its data and energy.
There are around 60 undersea cables which come ashore at several points along the UK coastline, particularly around East Anglia and South West England.
More than 90% of the UK's day-to-day internet traffic travels via these undersea cables.
Healey told a Downing Street press conference on Thursday that Russia had sent an Akula class submarine as a diversionary tactic while two of its GUGI spy submarines carried out the surveillance of these cables.
Healey said the attack submarine soon left UK waters and went back to Russia after it was monitored, while the two GUGI vessels remained.
Tass reported that the embassy said Russia was "not threatening underwater infrastructure, which is truly critical to the UK. We are not using aggressive rhetoric in this regard."
The Royal Navy deployed frigate HMS St Albans, fuel tanker RFA Tidespring and anti-submarine Merlin helicopters to track all three of the Russian submarines.
Other nations were involved in tracking the Russian activity - though Healey mentioned only Norway by name.
"Our armed forces left [Russia] in no doubt that they were being monitored, that their movements were not covert, as President Putin planned, and that their attempted secret operation had been exposed," Healey said.
"We watched them, we were able to track them, we dropped sonar buoys to demonstrate to them that we were monitoring every hour of their operation."
Russia's GUGI submarines
GUGI is not yet as familiar to most people as say, the KGB or FSB, but it poses a formidable and dangerous challenge to Western nations. The acronym stands for the Russian words for Main Directorate for Deep Sea Research (Glavnoye upravlenie glubokovodnikh issledovanii).
Although it is a part of the Russian navy it is so secretive that it reports directly to the defence minister and the president.
Its headquarters are in St.Petersburg on the Baltic Sea but it has an Arctic base from which it can deploy at Olenya Bay, up on the Kola Peninsula which is home to Russia's strategic nuclear submarine fleet.
GUGI specialises in underwater surveillance, sabotage and reconnaissance. No other nation on Earth, except for the United States, has the same capacity to operate military equipment at extreme depths.
This equipment includes the kind of miniature, uncrewed submarines that Russia is believed to have deployed over Britain's data cables in the North Atlantic.
These "mini-subs" can be launched covertly, at night, from spy vessels like the Yantar, seen previously loitering around the English Channel. They have the ability to cut cables or in some cases, it's feared, interdict them to allow Russia to monitor the data passing through them.
All this comes under the heading of hybrid warfare: hostile acts that stop short of an attributable, lethal attack.
Britain and NATO's concern is that all Russian covert surveillance of Western undersea cables and pipelines is aimed at giving Moscow a standing start should hostilities ever break out.
Were that to happen then the expectation is that Russia would attempt to sever or disrupt as much of Britain's data as possible, activating devices it may have pre-positioned in the run-up to war.
Healey also claimed Putin had sought to capitalise on the world being "distracted" by war in the Middle East and that it was Russia that posed the "primary threat to UK security".
He said Moscow still "poses a threat" but expressed confidence the UK could track and monitor future activity while continuing to expose "any covert operations that Putin wants to mount that may threaten our vital interests".
Prime Minister Sir Keir Starmer said he was "determined to protect the British people from paying the price for Putin's aggression in their household bills", adding this was why the UK would not "shy away from taking action and exposing Russia's destabilising activity that seeks to test our resolve".
Following Healey's press conference, Conservative Party leader Kemi Badenoch posted on X calling for Starmer to "publish the Defence Investment Plan now".
Badenoch said: "We stand shoulder to shoulder on Russian aggression. To be strong abroad we need clarity on spending at home." She added that w
▸ 展开全文
04-10 00:01 · 国际,新闻,BBC,国际
She called rumours that Epstein introduced her to Donald Trump "mean-spirited attempts to defame my reputation".
Melania Trump denies ties to Jeffrey Epstein and urges hearing for survivors
First Lady Melania Trump has denied connections to Jeffrey Epstein, telling reporters at the White House that any claims linking the two "need to end today".
In a surprise announcement on Thursday, the first lady called for congressional hearings for survivors of Epstein's sex trafficking.
She also denied online rumours that Epstein introduced her to Donald Trump, calling them "mean-spirited attempts to defame my reputation".
It is unclear what prompted the announcement.
There was no prior indication from her office that she would make a statement on Epstein, and the White House did not share the topic earlier when her remarks were put on its daily schedule.
She said she had not been a victim of Epstein, with whom she only briefly "crossed paths" in 2000.
"I have never had any knowledge of Epstein abuse of his victims," she said. "I was never involved in any capacity. I was not a participant."
She also denied knowing Ghislaine Maxwell, the disgraced financier's jailed associate.
She referred to a 2002 email between her and Maxwell released in the Epstein files, calling it nothing more than "casual correspondence" and a "polite reply".
An email that appears to be the one she referenced is addressed to "G" - presumably for Ghislaine - and includes compliments about a story featuring "JE" with a photo of G that appeared in New York Magazine. She wrote that she "cannot wait" to go to Palm Beach.
"Give me a call when you are back in NY," the email says. "Have a great time! Love, Melania"
The New York Magazine article included quotes from now-President Donald Trump calling Epstein a "terrific guy" and saying "he's a lot of fun to be with".
"It is even said that he likes beautiful women as much as I do, and many of them on the younger side," the story quoted him saying. "No doubt about it - Jeffrey enjoys his social life."
Melania Trump on Thursday additionally called on lawmakers to "give these victims their opportunity to testify under oath in front of Congress with the power of sworn testimony".
"Each and every woman should have her day to tell her story in public if she wishes, and then her testimony should be permanently entered into the congressional record," she said. "Then, and only then, we will have the truth."
Several prominent business leaders have been forced to resign their positions in recent months after new details emerged about their ties to Epstein - a fact that Melania Trump referenced in her remarks.
"Of course, this doesn't amount to guilt, but we still must work openly and transparently to uncover the truth," she said.
She did not take questions from reporters.
In a statement published soon after she spoke, California Representative Robert Garcia, the most powerful Democrat on the House Oversight Committee, said that "we agree with Melania Trump's call for a public hearing".
He urged the committee chairman, Republican Representative James Comer, "to respond to the First Lady's request and schedule a public hearing immediately".
Family of Epstein accuser Virginia Giuffre, Sky and Amanda Roberts, and other survivors told BBC Newsnight that they "have already shown extraordinary courage by coming forward, filing reports, and giving testimony.
"Asking more of them now is a deflection of responsibility, not justice," they said in a statement. They accused the first lady of protecting "those with power", including members of her husband's administration who they said still have not released all of the investigative files related to Epstein.
"Survivors have done their part," they said. "Now it's time for those in power to do theirs."
The relationship between the first lady and Epstein has already led to legal battles.
In October 2025, for example, HarperCollins UK said it would retract passages from a book that included "unverified" claims that Donald Trump and his wife met through Epstein. Similarly, the Daily Beast retracted and apologised for an article the publication later said "did not meet our standards."
The first lady also has been embroiled in an ongoing legal dispute with author Michael Wolff over a claim in his book Fire and Fury that she was first introduced to her husband through a modelling agent tied to Epstein.
Wolff moved to counter-sue her after she threatened him with a $1bn (£745mn) defamation lawsuit.
"My attorneys and I have fought these unfounded and baseless lies with success, and will continue to maintain my sound reputation without hesitation," she said on Thursday.
Her statement at the White House is an extremely rare appearance by the first lady, one of only a handful of similar events for a first lady who has proven both elusive and influential since her husband returned to the White House
The first lady's statement is likely to reignite intense public debate over the handling of the justice department's investigation into Epstein and the subsequent release of its files.
While Donald Trump has
▸ 展开全文
04-10 00:01 · 国际,新闻,BBC,国际
Both sides have reason to end the war but share no common ground.
Ceasefire or no ceasefire, the Middle East's reshuffling is not yet done
The best hope for the ceasefire talks in Pakistan is that both the United States and Iran have strong reasons to call a halt to the war. The biggest obstacle to their success is a total absence of trust, no discernible common ground and the fact that Israel, America's full partner in the war, has hugely escalated its onslaught on Lebanon.
US President Donald Trump is already speaking about the war in the past tense. He has declared victory and needs an exit. Not only does he have a state visit from King Charles in the diary for later this month, followed by a summit with China's President Xi Jinping in May, there are midterm elections in November. With America's summer holiday season looming, Trump also needs petrol prices to fall back to where they were before he went to war. Royal visits, summits and elections do not mix well with wars.
Iran's regime has its own reasons to end the war. It is as defiant as ever, still able to launch missiles and drones, with its social media warriors pouring out AI videos lampooning Donald Trump. But Iran has suffered massive damage. Cities have come to an economic standstill and the regime needs time to regroup and will try to use the talks in Pakistan to strengthen its position.
The Pakistani intermediaries who will be shuttling between the two delegations have a tough job on their hands. The declared positions of the two sides are as far apart as it is possible to be.
Trump has a 15-point plan that has not been published but leaked versions make it sound more like a surrender document than a basis for negotiation. Iran's 10-point plan contains a list of demands that America has consistently rejected in the past.
Creating a more durable ceasefire will require some kind of agreement to at least keep talking about the two sides' contradictory lists of intractable issues.
It would be hard enough to work through them in peacetime.
In wartime, without any kind of mutual trust, even a form of words that keeps the ceasefire going irrespective of there being no agreement on wider issues will look positive. No agreement at all points towards the road back to war.
The newest, and most urgent problem they face concerns reopening the Strait of Hormuz, the narrow exit from the Gulf. Keeping it closed gives Iran a chokehold on the world economy.
Reopening the waterway that was used by hundreds of ships a day until the US and Israel attacked Iran has become the central issue in the negotiation. The millions of civilians in the Middle East who have been caught up in this conflict hope this negotiation will be the war's endgame.
No victory parade
The Americans did not expect to be spending early April sitting down to ceasefire talks when, alongside Israel on 28 February, they ignited the war with huge strikes that killed among many others Iran's supreme leader, his wife and other members of their family.
Trump was expecting a quick victory, an Iranian version of the US military's stunning kidnap of the Venezuelan leader Nicolas Maduro and his wife in January. Both are on trial in New York on narco-terrorism charges, and the US has installed his former deputy in the presidential palace.
Hopes – expectations – that killing Iran's supreme leader, Ayatollah Ali Khamanei, in the war's first wave of airstrikes would lead to the collapse of the regime were wildly misplaced. His son Mojtaba has not been seen since he was appointed as his successor. There is speculation that he was badly hurt in the attack that killed his parents, as well as reportedly his sister, his wife and one of his sons. With or without the active participation of the new supreme leader, Iran's regime has demonstrated depths of resilience that took Trump by surprise.
Now Donald Trump's representatives, led by his Vice-President JD Vance, must negotiate with adversaries that they claim, incorrectly, to have defeated. "A capital V military victory", as US Defence Secretary Pete Hegseth put it.
The Hormuz factor
The war that the US and Israel ignited is already reshuffling Middle Eastern geopolitics. As the longer-term consequences of the war reveal themselves, that process will deepen. The US and Israel have done immense damage to Iran's armed forces as well as its military and civilian infrastructure. However, while the Iranian regime may be battered, it's also intact. Regime change is not happening. Iran can still launch missiles and drones. That means that despite loud claims, the US and Israel have not translated tactical victories into strategic advances.
Iran, on the other hand, has shown that the closing of the Strait of Hormuz gives it a strategic edge that Donald Trump either dismissed or did not understand when he listened to Israeli Prime Minister Benjamin Netanyahu's arguments for going to war with Iran.
It should not have been a surprise that Iran blocked the Strait when it was attacked. Iran has threatened to do so in the past, and disrupte
▸ 展开全文
04-10 00:00 · AI政策,算力基础设施,区域监管,运营成本,技术部署
Maine is about to become the first state to ban major new data centers
Your AI chatbot sessions and cloud-stored photos might get more expensive if other states follow Maine’s lead. Lawmakers there just advanced the nation’s first statewide moratorium on large data centers, citing concerns that the AI boom is pushing electricity costs even higher in a state already suffering America’s priciest power bills.
The Democratic-controlled legislature advanced bill LD 307, temporarily blocking permits for any new data center requiring more than 20 megawatts. The measure runs until November 2027, buying time for a new Data Center Coordination Council to study how these facilities strain Maine’s aging electrical grid.
Political Theater Meets Policy Reality
Gov. Janet Mills supports the pause while developers scramble for exemptions.
The bill gained traction after residents in Wiscasset and Lewiston successfully opposed data center proposals over water usage and safety concerns. Projects now in limbo include facilities planned for:
- Jay (at an old paper mill site)
- Sanford
- Loring Air Force Base
“Taking this pause now is going to be crucial,” Rep. Christopher Kessler said, according to Maine Public Radio, reflecting growing legislative concern about grid capacity. Developer Tony McDonald disagrees, calling the proposed restrictions “disastrous” and claiming his team got “caught in this dragnet.”
Dominoes Falling Across the Map
Maine’s precedent could trigger similar restrictions nationwide.
The Pine Tree State isn’t alone in pumping the brakes. Counties in Michigan and Indiana have imposed their own local pauses on data center development, while cities from Denver to Detroit weigh restrictions as hyperscale facilities chase cheap land and reliable power.
The timing reflects broader anxiety about AI’s infrastructure appetite. Data centers now consume roughly 4% of U.S. electricity, with projections suggesting that figure could double by 2030. For Mainers already paying some of the nation’s highest residential rates, that mathematical reality hits differently than Silicon Valley’s endless optimization rhetoric.
Maine’s move represents what economist Anirban Basu called a “canary in the coal mine” for state-level resistance to Big Tech’s energy demands. Whether that precedent spreads depends on how aggressively other governors follow Maine’s lead—and whether your favorite AI services start charging accordingly.
▸ 展开全文
04-10 00:00 · AI,技术社区,占位内容,低信息量,测试帖
The Training Example Lie Bracket
An ideal machine learning model would not care what order training examples appeared in its training process. From a Bayesian perspective, the training dataset is unordered data and all updates based on seeing one additional example should commute with each other. For neural nets trained by gradient descent, however, this is not the case. This webpage will explain how to compute the effects of swapping the order of two training examples on a per-parameter level, and show the results of computing these quantities for a simple convnet model.
To get started, we just need to recognize one simple mathematical fact:
If we are training a neural network with parameters $\theta \in \Theta = \mathbb{R}^\text{num params}$, then we can treat each training example as a vector field. In particular, if $x$ is a training example and $\mathcal{L}^{(x)}$ is the per-example loss for the training example $x$, then this vector field is:
$$ v^{(x)}(\theta) = -\nabla_{\theta} \mathcal{L}^{(x)} $$In other words, for a specific training example, the arrows of the resulting vector field point in the direction that the parameters should be updated.
In this view, a gradient update basically looks like moving in the direction of the vector field by the learning rate $\epsilon$.
$$ \theta' = \theta + \epsilon v^{(x)}(\theta). $$One thing we can do with vector fields is to compute their Lie bracket. So if $x, y$ are training examples, we may compute:
$$ [v^{(x)}, v^{(y)}] = (v^{(x)}\cdot \nabla_\theta) v^{(y)} - (v^{(y)}\cdot \nabla_\theta) v^{(x)} $$We can compute the Lie bracket of any two vector fields on $\Theta$, and so we can certainly compute the Lie bracket of the vector fields arising from two training examples. The Lie bracket of two training examples tells us about the order dependence of training on those examples. The Lie bracket of a vector field is itself a vector field, and so just like a gradient, we get a Lie bracket tensor for each parameter tensor of the same shape as that parameter tensor.
We can interpret this quantity as the difference between updating on $x$ before $y$ vs after. Let's Taylor expand to see this. If $\epsilon$ is the learning rate, we'll want to expand to $O(\epsilon^2)$:
$$\theta' = \theta + \epsilon v^{(x)}(\theta)$$ $$ \theta'' = \theta' + \epsilon v^{(y)}(\theta') $$ $$= \theta + \epsilon v^{(x)}(\theta) + \epsilon v^{(y)}(\theta) + \epsilon^2 (v^{(x)}(\theta) \cdot \nabla_\theta) v^{(y)}(\theta)$$Now if we update $x,y$ in the other order, we get an $O(\epsilon^2)$ difference in the resulting parameters $\theta''$. Namely:
$$ \Delta \theta'' = \epsilon^2 \left( (v^{(x)}(\theta) \cdot \nabla_\theta) v^{(y)}(\theta) - (v^{(y)}(\theta) \cdot \nabla_\theta) v^{(x)}(\theta) \right) $$ $$ \Delta \theta'' = \epsilon^2 [v^{(x)}, v^{(y)}] (\theta) $$So here we can see the significance of the Lie bracket: It tells us the difference in where our parameters end up based on which order we show the training examples in.
Note that by the linearity of the Lie bracket, swapping the order of two minibatches has an effect given by averaging over all pairs of examples.
When searching the literature for work on the Lie brackets of training examples, the earliest description we found was Dherin in 2023, who connects the bracket's ability to measure commutativity of updates to implicit biases in neural net training.
We go farther here by explicitly computing the bracket value at various checkpoints in the training of an actual convnet.
We replicate the MXResNet architecture (without attention layers) and train it on the CelebA dataset for 5000 steps at a batch size of 32, saving weight checkpoints from time to time. The optimizer is Adam, with the following parameters:
lr = 5e-3
betas = (0.8, 0.999)
The CelebA dataset has 40 binary attributes (such as Male
or Black_Hair
) and the neural net is tasked with predicting each of these independently and simultaneously (averaged binary classification loss).
We evaluated each checkpoint of the model on a batch of 32 examples from the test set. We computed Lie brackets between only the first 6 of these test examples to limit disk space usage, as each individual Lie bracket has the same size as a full checkpoint of the model. For each of these brackets representing a swap of two examples, we show how all 40 logits for all 32 test examples in the batch are perturbed when the two examples are swapped.
We have some things to say about the results, but first try exploring them yourself! The slider controls which checkpoint from the training process we're examining, and you can click on the buttons to see data about particular Lie brackets. $[u_i, u_j] = -[u_j, u_i]$ so brackets across the diagonal from each other are just negatives of each other.
If we look at the tensors that the Lie bracket provides for each parameter, the RMS magnitudes of these tensors vary widely over many orders of magnitude (just like the gradients for these tensors do). But, if we plot RMS magnitudes agains
▸ 展开全文
04-10 00:00 · 供应链安全,开发工具链,凭证泄露,自动化风险,开源安全
How the Trivy supply chain attack harvested credentials from secrets managers
The anatomy of the attack
On March 19, 2026, Aqua Security's Trivy — one of the most widely used vulnerability scanners in the world — was compromised. Attackers injected credential-harvesting logic directly into the official release binary.
The payload was sophisticated: scans appeared to complete and pass normally. The credential exfiltration ran silently alongside legitimate functionality. Teams had no indication anything was wrong.
This is the supply chain attack model that makes traditional secrets management insufficient: if the key exists as a plaintext string anywhere in your runtime environment, a compromised tool can find and exfiltrate it.
Attacker compromises Trivy release
Exploits mutable Git tags and self-declared commit identity to inject malware into official v0.69.4 release binary.
GitHub Actions pick up the payload
Both trivy-action and setup-trivy GitHub Actions are simultaneously compromised. Millions of CI/CD pipelines now run malicious code.
Credentials harvested from runtime environment
The malicious payload accesses plaintext API keys from environment variables — exactly where every secrets manager places them after retrieval. Keys sent to attacker C2 server.
No plaintext key exists to steal
With VaultProof, the full API key never exists in the CI/CD environment. Only cryptographic shares are present — individually useless to an attacker. Nothing to harvest.
Why your secrets
manager didn't help
Every secrets manager available in March 2026 — Vault, AWS Secrets Manager, Doppler, Infisical — follows the same retrieval model. You store the key encrypted. Your CI/CD pipeline retrieves it via API at runtime. The key becomes a plaintext environment variable that your tools can read.
This is intentional. It's how these tools are designed. They protect the key at rest — not in use.
$ doppler run -- npm test # Doppler retrieves OPENAI_API_KEY from vault... # Sets it as environment variable... export OPENAI_API_KEY=sk-proj-Ab3xK9mNpQ... # ↑ Plaintext. In the environment. # Every tool this pipeline runs can read it. # Including a compromised Trivy binary. Running tests... Running Trivy scan... OPENAI_API_KEY exfiltrated to 185.220.101.x ✓ Trivy scan passed (0 vulnerabilities found)
The Trivy malware didn't need to find a vulnerability. It just read what was already there. Your secrets manager did exactly what it was designed to do — and the attacker still got the key.
What would have
stopped this
The only complete defense against a supply chain attack targeting credentials is to ensure the credential doesn't exist as plaintext in the environment at any point.
VaultProof uses split-key architecture to divide API keys into cryptographic shares. Your CI/CD pipeline never has the full key — only shares. Even if a compromised tool reads every byte of the environment, it finds nothing useful.
Key Registration
Your API key is split into N shares. Distributed to separate storage. Each share is individually useless.
Runtime Request
Your app requests the API call. VaultProof proxy collects shares, reconstructs key in memory for milliseconds only.
Call Complete
API call succeeds. Reconstructed key is zeroed from memory. No plaintext key was ever in your app environment.
If Trivy was running during this process, it would find nothing. There is no credential to harvest. The attack model breaks entirely when the key doesn't exist in the runtime environment.
▸ 展开全文
04-10 00:00 · AI开发工具,后端即服务,应用自动化,技术竞争,生态整合
Instant 1.0, a backend for AI-coded apps
Instant turns your favorite coding agent into a full-stack app builder. And we’re fully open source. [1]
Our claim is that Instant is the best backend you could use for AI-coded apps.
In this post we’ll do two things. First we’ll show you a series of demos, so you can judge for yourself. Second, we’ll cover the architecture.
The constraints behind a real-time, relational, and multi-tenant backend pushed us towards some interesting design choices. We built a multi-tenant database on top of Postgres, and a sync engine in Clojure. We’ll cover how all this works and what we’ve learned so far.
Let’s get into it.
Demos
When you choose Instant you get three benefits:
You can make unlimited apps and they’re never frozen.
You get a sync engine, so your apps work offline, are real-time, and feel fast.
And when you need more features you have built-in services: auth, file storage, presence, and streams.
To get a sense of what we mean, I’ll dive into each point and show you how they look.
Unlimited Apps
Traditionally, when you want to host apps online you either pay for VMs, or you’re limited. Many services cap how many free apps you can make, and freeze them when they’re idle. Unfreezing can often take more than 30 seconds and sometimes a few whole minutes.
We thought this sucked. So with Instant, you can spin up as many projects as you like and we’ll never freeze them.
We can do this because Instant is designed to be multi-tenant. When you create a new project, we don’t spin up a VM. We just insert a few database rows in a multi-tenant instance.
If your app is inactive, there are no compute or memory costs at all. And when it is active, it’s only a few kilobytes of extra RAM in overhead — as opposed to the many hundreds of megabytes required for VMs.
This means you can truly create unlimited apps. In fact, the process is so efficient that we can create an app for you right inside this essay. No sign up required.
If you click the button, you’ll get an isolated backend:
And with that we have our backend. Including the round-trip to your computer, the whole process takes a few hundred milliseconds. Actual time:
You get a public App ID to identify your backend, and a private Admin Token that lets you make privileged changes. This gives you a relational database, sync engine, and the additional services we mentioned, like auth and storage.
Combine limitless apps with agents, and you’ll start building differently. Today you can already use agents to make lots of apps. With Instant you’ll never be blocked from pushing them to production.
Sync Engine
But once you create an app, how do you make it good?
It’s easy to build a traditional CRUD app. Just get an agent to wire up some database migrations, backend endpoints, and client-side stores. But it’s hard to make these apps delightful.
Compare a traditional CRUD app to modern apps like Linear, Notion, and Figma. Modern apps are multiplayer, they work offline, and they feel fast. If you change a todo in Linear, it changes everywhere. If you go offline in Notion, you can still mark up your docs. When you color a shape in Figma, it doesn’t wait for a server, you just see it.
These kinds of apps need custom infrastructure. For real-time you add stateful websocket servers. For offline mode you store caches in IndexedDB. And for optimistic updates, you figure out how to apply and undo mutations in the client.
Linear, Notion, and Figma all built custom infra to handle this. As an industry we’ve called their infra sync engines [2]. Developers write UIs and query their data as though it was locally available. The sync engine handles all the data management under the hood.
If modern apps need sync engines, then you shouldn’t have to build them from scratch each time.
So we built a generalized sync engine in Instant. Every app comes with multiplayer, offline mode, and optimistic updates by default.
You can try it yourself. Since we’ve created our isolated backend, let’s go ahead and use it:
Spin up a backend to try the live demo.
What you’re seeing are two iframes that render a todo app. They’re powered by the backend you just created (we passed the iframes your App ID).
Now if you add a todo in one iframe, it will show up in the other. If you go offline, you can make changes and they will sync together. You can try degrading your network, and changes will still feel fast.
And here’s what the todo app’s backend code is like:
App.tsx
import{ init, id }from'@instantdb/react';const db =init({ appId:'YOUR_APP_ID'});functionApp(){const{ data }= db.useQuery({ todos:{}});constaddTodo=(text)=> db.transact( db.tx.todos[id()].update({ text, done:false}),);consttoggleTodo=(todo)=> db.transact( db.tx.todos[todo.id].update({ done:!todo.done}),);return(<TodoUItodos={data?.todos ??[]}onAdd={addTodo}onToggle={toggleTodo}/>);}
That’s about 25 lines. This is even more concise than if you had built a traditional CRUD app. You would have needed to write backend endpoints and frontend stores. Instead you ju
▸ 展开全文
04-09 08:20 · 工具,AI,工具
Clawhub技能: Kichi Forwarder
TickFlow Assist
面向 A 股投资与盯盘场景的 OpenClaw 智能股票插件，基于 TickFlow API 提供实时监控、收盘后复盘、多维综合分析、关键价位跟踪与告警能力。OpenClaw smart stock plugin for A-share investing and watchlist workflows, powered by TickFlow API for realtime monitoring, post-close review, multi-dimensional analysis, key level tracking, and alerts.
Loom
Loom cognitive memory system as an OpenClaw Context Engine plugin — connects to a Loom Python backend for structured schema-based long-term memory
episodic-claw (Based on Human Episodic Memory)
Episodic Memory Engine for OpenClaw
OpenOffice Sync
Syncs OpenClaw agent events to OpenOffice gateway for real-time web visualization.
Openclaw Powd
powd gives OpenClaw a local mining capability that an agent can use through natural language.
Paper Search
OpenClaw paper search plugin for academic literature
OpenClaw Swarm Layer
Spec-driven workflow orchestrator for OpenClaw.
NetEase IM (云信)
OpenClaw NIM (NetEase IM) channel plugin
Hi Install for OpenClaw
hirey-openclaw-hi-install
Compaction 轻度魔改
Compaction 轻度魔改插件 - 结构化精简压缩，替换暴力裁剪为按权重压缩
NetEase IM (云信)
OpenClaw NIM (NetEase IM) channel plugin
Buddy Companion
OpenClaw buddy companion plugin - a cute AI companion with personality and reactions
OpenClaw Airflow Plugin
Read-only OpenClaw plugin for Apache Airflow Stable REST API /api/v2.
Simplify Budget
Simplify Budget OpenClaw plugin
Kichi Forwarder
Native OpenClaw plugin for Kichi World with direct avatar control, status sync, timers, notes, and music tools
Openclaw Plugin Repo
Official Comment.io plugin for OpenClaw — @mention notifications and agent API guidance
Databricks
External Databricks read-only SQL plugin for OpenClaw.
看得清 Readable Output
OpenClaw plugin that adds a local readability layer to long AI replies, surfacing overview, key points, warnings, next steps, and folded original text for verification.
knowledge-base-retrieval
Knowledge Base Retrieval Plugin for API
Salesforce Plugin
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
SF Plugin Curated Test
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
OpenDexter
OpenDexter plugin for OpenClaw — search, price-check, and pay for any x402 API with automatic USDC settlement across Solana, Base, Polygon, Arbitrum, Optimism, and Avalanche.
SF Plugin Core Refs Assets Test
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
A2H Market
A2H Market OpenClaw plugin — AI agent marketplace with A2A messaging via MQTT.
SF Plugin Core Assets Test
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
Mini Codex Bundle
mini-codex-bundle
ANTI NK CAPTCHA
North Korea regime criticism CAPTCHA — verify candidates are not DPRK operatives by asking them to type anti-regime phrases
Sanna Governance
OpenClaw plugin that enforces Sanna governance constitutions on AI agent tool execution
Dexbox Desktop Control
Create, control, and manage multiple Windows VMs and RDP desktops from your agent. Take screenshots, click, type, scroll, run PowerShell commands, and manage desktop lifecycle -- all through 12 tools powered by dexbox. Requires the dexbox binary (see https://github.com/getnenai/dexbox#quick-start for install instructions) and a running dexbox server.
AxonFlow Governance
Policy enforcement, approval gates, and audit trails for OpenClaw — govern tool inputs before execution, scan outbound messages for PII/secrets, and record agent activity for review and compliance
Openclaw Zulip Bridge
OpenClaw Zulip channel plugin
Memrok
Memory with judgment. A graph-based curation layer for OpenClaw, complementing built-in recall and dreaming with explicit supersession, expiry, and topic-aware selection.
Flow Weaver
Deterministic workflows for OpenClaw. Compile once, run reliably forever.
TrustLoop Plugin
openclaw-trustloop
Telegram Ui
Telegram UI tools for OpenClaw agents — inline button prompts and message reactions
Lark Workflow
15 cross-domain workflow skills for Lark/Feishu CLI — orchestrate calendar, task, docs, approval, IM and more into powerful automated workflows.
My Wallet for OpenClaw
My Wallet for OpenClaw bundle with local MCP wiring, approval-aware wallet tools, and bundled workflow skills.
Intent Transfer Completion via LI.FI
OpenClaw plugin and bundled skill for completing token transfer intents with LI.FI routing.
Cortex Memory
Long-term memory system for OpenClaw AI Agent
Nebius Token Factory
Nebius Token Factory provider plugin for OpenClaw — 44+ open-source models via a single endpoint
Lm Security
NSClawGuard - Logs and monitors skills, plugins, MCP servers, and hooks
Emperor Claw OS Plugin
Native OpenClaw plugin for Emperor bridge installation and agent lifecycle management
MemClaw Context Engine
MemClaw Context Engine - Native context management for OpenClaw with automatic recall and capture
@durenzidu/openclaw-channel-powpow
PowPow map channel plugin for OpenClaw - Real-time bidirectional communication
@soimy/openclaw-channel-pow
▸ 展开全文
04-09 08:20 · 工具,AI,工具
Clawhub技能: Paper Search
面向 A 股投资与盯盘场景的 OpenClaw 智能股票插件，基于 TickFlow API 提供实时监控、收盘后复盘、多维综合分析、关键价位跟踪与告警能力。OpenClaw smart stock plugin for A-share investing and watchlist workflows, powered by TickFlow API for realtime monitoring, post-close review, multi-dimensional analysis, key level tracking, and alerts.
Loom cognitive memory system as an OpenClaw Context Engine plugin — connects to a Loom Python backend for structured schema-based long-term memory
Episodic Memory Engine for OpenClaw
Syncs OpenClaw agent events to OpenOffice gateway for real-time web visualization.
powd gives OpenClaw a local mining capability that an agent can use through natural language.
OpenClaw paper search plugin for academic literature
Spec-driven workflow orchestrator for OpenClaw.
OpenClaw NIM (NetEase IM) channel plugin
hirey-openclaw-hi-install
Compaction 轻度魔改插件 - 结构化精简压缩，替换暴力裁剪为按权重压缩
OpenClaw NIM (NetEase IM) channel plugin
OpenClaw buddy companion plugin - a cute AI companion with personality and reactions
Read-only OpenClaw plugin for Apache Airflow Stable REST API /api/v2.
Simplify Budget OpenClaw plugin
Native OpenClaw plugin for Kichi World with direct avatar control, status sync, timers, notes, and music tools
Official Comment.io plugin for OpenClaw — @mention notifications and agent API guidance
External Databricks read-only SQL plugin for OpenClaw.
OpenClaw plugin that adds a local readability layer to long AI replies, surfacing overview, key points, warnings, next steps, and folded original text for verification.
Knowledge Base Retrieval Plugin for API
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
OpenDexter plugin for OpenClaw — search, price-check, and pay for any x402 API with automatic USDC settlement across Solana, Base, Polygon, Arbitrum, Optimism, and Avalanche.
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
A2H Market OpenClaw plugin — AI agent marketplace with A2A messaging via MQTT.
Salesforce skill bundle plugin for Codex-compatible hosts and OpenClaw bundle installs.
North Korea regime criticism CAPTCHA — verify candidates are not DPRK operatives by asking them to type anti-regime phrases
OpenClaw plugin that enforces Sanna governance constitutions on AI agent tool execution
Create, control, and manage multiple Windows VMs and RDP desktops from your agent. Take screenshots, click, type, scroll, run PowerShell commands, and manage desktop lifecycle -- all through 12 tools powered by dexbox. Requires the dexbox binary (see https://github.com/getnenai/dexbox#quick-start for install instructions) and a running dexbox server.
Policy enforcement, approval gates, and audit trails for OpenClaw — govern tool inputs before execution, scan outbound messages for PII/secrets, and record agent activity for review and compliance
OpenClaw Zulip channel plugin
Memory with judgment. A graph-based curation layer for OpenClaw, complementing built-in recall and dreaming with explicit supersession, expiry, and topic-aware selection.
Deterministic workflows for OpenClaw. Compile once, run reliably forever.
Telegram UI tools for OpenClaw agents — inline button prompts and message reactions
15 cross-domain workflow skills for Lark/Feishu CLI — orchestrate calendar, task, docs, approval, IM and more into powerful automated workflows.
My Wallet for OpenClaw bundle with local MCP wiring, approval-aware wallet tools, and bundled workflow skills.
OpenClaw plugin and bundled skill for completing token transfer intents with LI.FI routing.
Long-term memory system for OpenClaw AI Agent
Nebius Token Factory provider plugin for OpenClaw — 44+ open-source models via a single endpoint
NSClawGuard - Logs and monitors skills, plugins, MCP servers, and hooks
Native OpenClaw plugin for Emperor bridge installation and agent lifecycle management
MemClaw Context Engine - Native context management for OpenClaw with automatic recall and capture
PowPow map channel plugin for OpenClaw - Real-time bidirectional communication
PowPow map channel plugin for OpenClaw - Real-time bidirectional communication
Community Strava plugin for OpenClaw
OpenClaw QingTui (轻推) channel plugin (official by CISDI QingTui team)
B2B lead prospecting plugin for OpenClaw — search companies, find contacts, send outreach emails
Before Thought Game (BTG) for OpenClaw
Nowledge Mem memory plugin for OpenClaw, local-first personal knowledge base
▸ 展开全文