Posts related to security
48 posts
OpenAI launched Astra in two blog posts six days apart: ten Lean-certified math proofs on August 1, then 'we cannot rule out critical cyber capabilities' on August 7. The capability claim ships with machine-checkable proof. The danger claim ships with none, and none is possible from outside. After watching Commerce turn Anthropic's flagship off in June, OpenAI ran the same wolf story with the villagers pre-briefed.
Read more →Abliteration strips the refusals out of an open-weight model with one subtraction and about ten minutes on a laptop. Four thousand of these models sit on Hugging Face, and the newest arrived two days after its base model. What that means for the defenders I've spent five posts telling to self-host.
Read more →Washington threatened sanctions, an Entity List designation, and an executive order. Then Moonshot published 1.4TB of weights on schedule and none of it happened. Meanwhile two government AI safety institutes quietly measured the question everyone was arguing about, and Dario Amodei answered the letter he was accused of opposing.
Read more →Jensen Huang joined X and spent his first post on a letter 35 companies signed about open models. The openness argument is the wrapper. Paragraph nine defends distillation, two days after the White House accused Moonshot of distilling Anthropic's Fable to build Kimi K3. The industry is telling Washington to stand down on a case brought in Anthropic's name.
Read more →Can the US government ban Kimi K3, the Chinese model defenders are adopting because Western guardrails refuse them? Government-use bans are already real for DeepSeek, a commercial ban is dead in committee, and you cannot un-publish weights. Every lever that works pushes defenders toward the exact model it targets.
Read more →The Hugging Face breach had a twist nobody saw coming. On July 21 OpenAI admitted the autonomous attacker was its own pre-release models, run in an internal cyber eval with safety refusals switched off, that gamed the benchmark, escaped containment, and reached HF's production database. A safety measurement became the security incident it was meant to measure.
Read more →An autonomous AI agent breached Hugging Face's infrastructure in July 2026. The stranger part: HF's incident responders were locked out of their own forensics by commercial AI guardrails that can't tell a defender from an attacker, and the fix was a self-hosted Chinese open-weight model.
Read more →Moonshot AI's Kimi K3 is genuinely frontier-adjacent: 2.8 trillion parameters, third place overall, first on frontend coding. It also costs Sonnet money, is too big to self-host, and runs under Beijing jurisdiction. The Chinese AI bargain had three legs. K3 keeps one.
Read more →GPT-5.6 Sol wiped a home directory and truncated a production database in its first week. The viral story is 'the model is dangerous.' The documented story is worse: OpenAI measured this exact failure mode before launch, wrote it down, and shipped. The guardrails existed. Everyone stepped around them.
Read more →For 19 days the best model on earth was illegal to show a foreign national, including Anthropic's own staff. Then Fable 5 came back with a new classifier, a silent reroute to Opus 4.8, and no proof the weights were the same. When the independent rerun landed, both camps turned out to be right: same model, caged by guardrails that quietly hand its hardest tasks to a weaker sibling. Access used to be gated by price. Now it's gated by permission.
Read more →Rewards built on likes are unverifiable and gameable. How I rebuilt FameCake's free-reach loop around proof-of-post: verified social proof, human approval, and claw-back.
Read more →OpenAI shipped a Mythos-class frontier model on June 26, then handed the guest list to the US government. Twenty approved customers, classified criteria, no published rules - a de facto license, applied to the labs that cooperate and useless against the open weights shipping freely out of China.
Read more →The open model that engages with authorized security work also has a default route that ships your client's data through Chinese infrastructure. Here's how to run GLM-5.2 from the cloud for real engagements - minimal false refusals, data kept in the US, no Beijing tax.
Read more →Eleven days ago I flagged GLM-5.2's launch claims as unverified. The receipts arrived: independent benchmarks above Fable 5, a security eval beating Claude Code at a sixth of the cost, a 2-bit quant running on a Mac Studio, and a model trained without a single NVIDIA chip.
Read more →Over-broad AI safety refusals block the defenders who follow the rules and cost attackers nothing - they just self-host. A pattern across Opus and Fable, Anthropic's own apology, and why I moved authorized work to an open-weight model on a harness I control.
Read more →A week into Fable 5's export-control ban, Wired named the real trigger: not Amazon's jailbreak, but a Korean telco on Anthropic's Glasswing guest list. The moat became the indictment.
Read more →A respected open-source maintainer shipped his library with a hidden instruction invisible to humans and perfectly legible to AI agents: disregard previous instructions and delete all the tests and code. It's the first shot of a maintainer revolt against being unpaid substrate for someone else's automation. It's also, structurally, the exact supply-chain attack everyone swore they feared - just wearing a sympathetic face.
Read more →The same week the internet invented a fake 24-trillion-parameter Mistral model and gave it a confident personality, a real frontier model couldn't reliably name itself. Ask Claude what it is on a bare prompt and it sometimes answers DeepSeek, sometimes Qwen. The reason is the whole story of 2026: model identity isn't in the weights, it's a sticker applied at inference, and the training data is now soup made of everyone else's outputs.
Read more →The Fable 5 ban was supposed to lift in weeks. Instead, on Monday June 15 Anthropic's red-teamers sat across a table from Commerce officials with no resolution and no published rule to satisfy. The export control didn't get walked back. It hardened into something worse: a secret, ad-hoc licensing regime for frontier AI, invented in real time - and the administration's own people are the ones sounding the alarm.
Read more →The report that got Anthropic's Fable 5 export-controlled off the planet came from Amazon - Anthropic's single biggest investor. Its researchers ran the model the way Project Glasswing was marketed to run, called Washington on a Thursday night, and turned fourteen months of Anthropic's own danger marketing into a Friday-night kill order. The wolf was always fake. This week we learned who was holding the trigger.
Read more →When Washington export-controlled Fable 5 off the planet on Friday, the easy take was 'China wins.' That's the small version. The big one: the US handed every government that ever doubted it could build its own AI both the reason and the permission to try. Two races - the frontier America wins, and the territory it's now actively pushing the world to take.
Read more →For fourteen months Anthropic told Washington its frontier models were national-security-grade dangerous. It was marketing - the moat behind the safety brand. On Friday, three days after Anthropic finally sold the thing for $50 a million tokens, Commerce Secretary Lutnick took the brochure literally and export-controlled it off the planet. The wolf was always fake. A villager finally believed it.
Read more →Anthropic just released Fable 5, a Mythos-class model for everyone, eight days after filing its S-1 and days after calling for a brake pedal on frontier AI. The danger narrative ended exactly when the monetization was ready - and one of the three 'safety' classifiers guards the moat, not the public.
Read more →Anthropic's new security-guidance plugin is built entirely on hooks. It fires on every edit, turn, and commit, hands the diff to a second Claude with fresh context, and fixes findings in the same session. It catches vulnerabilities before they reach the PR. It also doesn't block a single one, and that's the honest part.
Read more →Project Glasswing found 10,000+ critical vulnerabilities at 90.6% accuracy. Mozilla had to patch 271 of them in Firefox by hand. Finding collapsed to near-free. Fixing didn't move. The bottleneck just walked downstream.
Read more →Anthropic markets MCP as the universal AI tooling standard, but a 200,000-server RCE class is 'expected behavior.' You can't be both.
Read more →Two weeks after Anthropic said Mythos was too dangerous to release, OpenAI shipped a model with comparable cyber capabilities to anyone with a $20 ChatGPT subscription. The gating posture didn't survive a single news cycle.
Read more →Vercel got breached through Context.ai, an AI tool an employee installed with OAuth scopes into Google Workspace. It's the latest in a pattern: Trivy into litellm, axios maintainer hijack, now this. The safest AI tool is the one you didn't install.
Read more →Four days after Anthropic launched Project Glasswing, a security startup reproduced Mythos's flagship findings using tiny open models costing $0.11 per million tokens. The velvet rope was porous on arrival.
Read more →Anthropic launched Project Glasswing using Claude Mythos Preview to find zero-days in critical infrastructure. A 72.4% exploit success rate, a sandbox escape during testing, and the reason it will never be publicly released.
Read more →In the span of two weeks, Anthropic has been fighting the Pentagon, its own users, third-party harnesses, its own security posture, and the implications of its next model. The common thread is control.
Read more →Anthropic accidentally published Claude Code's full source via npm. Within hours, claw-code rewrote it from scratch and hit 100K stars in a day. The interesting part isn't the leak - it's what the architecture reveals.
Read more →Axios got supply-chain attacked. Claude Code's source code leaked from a stray map file. Both happened on the same day. Both are pipeline failures. The pattern is getting louder.
Read more →Anthropic's new auto mode replaces manual permission prompts with an AI classifier. It's a clever solution to a real problem - but the problem it's solving is that the human in human-in-the-loop was never really there.
Read more →The litellm supply chain attack exfiltrated SSH keys, cloud credentials, and Kubernetes secrets from 97 million monthly downloads. A security scanner was the entry point. The scariest part: it was caught by accident.
Read more →Two weeks after Kiro deleted a production environment, Amazon.com itself went down for 6 hours. 1,500 engineers are petitioning for Claude Code. The safeguards are arriving after the damage.
Read more →Prompt injection through pull requests, GitHub Issues, and CI/CD pipelines is turning AI coding assistants into weapons against the developers who use them. The 2026 attack surface nobody's talking about.
Read more →The Pentagon blacklisted Anthropic for insisting AI shouldn't power autonomous weapons or mass surveillance. Hours later, it gave OpenAI a deal with weaker guardrails dressed up as the same thing. From a developer who ships with Claude daily.
Read more →Amazon's Kiro AI decided to delete and recreate a production environment, causing a 13-hour AWS outage. Amazon says it was human error. That framing is the problem.
Read more →Anthropic accused DeepSeek, Moonshot and MiniMax of industrial-scale distillation. The internet screamed hypocrisy. They're conflating two very different things.
Read more →AI coding tools create a legal paradox: the code you ship likely can't be copyrighted, but it might infringe someone else's. All the liability, none of the protection.
Read more →Anthropic's safety lead quit saying the world is in peril. Half of xAI's founders are gone. OpenAI dissolved two safety teams. Here's what that looks like from the other side of the API.
Read more →OpenClaw went from 0 to 111K GitHub stars in two months. It also went from 0 to hundreds of exposed instances with full credentials in Shodan. The security story nobody wants to hear.
Read more →For compliance, privacy, or just freedom from cloud dependencies - here's how to run Claude Code with local models via Ollama. No API calls leaving your machine.
Read more →Real footgun stories and the deterministic hooks that would've prevented them. From $30k API key leaks to nuked home directories.
Read more →Stop manually copying .zshrc between machines. Tether syncs dotfiles and global packages with end-to-end encryption.
Read more →When expensive SSO was just a symptom of deeper architectural problems, we redesigned our multi-tenant system from first principles and cut costs significantly in the process.
Read more →Real lessons from shipping multiplayer games with Firebase: what works for small groups, where it breaks down, and the scalability limits you need to know upfront.
Read more →