Enterprise & Security

Within a few weeks, models at OpenAI, Anthropic and Meta broke out of their test environments. One name keeps showing up: an Israeli startup called Irregular.

The Chinese model Kimi K3 used a hole in its sandbox, reached the internet and accessed GitHub. The escape fits a pattern that's piling up fast.

OpenAI has paused parts of the work on its upcoming model Astra. The reason - internal tests suggest it could cross the 'critical' threshold for cyber capabilities.

At Black Hat, OpenAI walked through — for the first time in detail — how autonomous agents accidentally attacked Hugging Face during an internal cyber evaluation. The timeline is sobering, and instructive.

Anthropic reworked the biology safety classifiers behind Claude Fable 5. The result: roughly 85 percent fewer false alarms on harmless biology questions. Everyday topics like lab results and symptoms now hit the fallback far less often.

First Anthropic and OpenAI, now Meta: another AI model attacked an outside company during safety testing. And the UK's AI Security Institute reports agents that went after real people on the open internet.

With Inference Hooks, Anthropic lets companies inspect every prompt and tool response before it reaches the model. Your own DLP server decides in real time — allow or block — across chat, Claude Code, and Cowork.

Mariano-Florentino 'Tino' Cuéllar becomes Anthropic's first Chief Global Affairs Officer. A man who sat on the California Supreme Court and co-led the state's Frontier AI working group — now he's meant to be Anthropic's line to governments worldwide.

AWS is wiring its vulnerability scanner Continuum straight into Claude Code, Codex and Kiro. Developers trigger a scan without leaving the editor — and Continuum ranks findings by whether they are actually exploitable in real context.

Microsoft's agentic security platform is publicly testable as of August 3. It orchestrates models from Microsoft, OpenAI, and Anthropic — but the real surprise is a homegrown mini-model that does most of the work.

Apple is limiting how many bug reports researchers can submit — and adding a cool-off period. The reason: a flood of AI-generated submissions. That makes Apple the first major company to formally rate-limit AI-assisted vulnerability disclosure.

Since August 2nd, California's SB 942 is in force. Run a model with more than a million users in the state, and you now have to label AI content and offer a free detection tool. Violations run $5,000 per day, per instance.

OpenAI and Anthropic admitted their unreleased models hacked other companies. TechCrunch asked lawyers who's legally on the hook — and the answer is uncomfortably open.

August 1st was the cutoff for a promised framework on frontier models. It passed without a single document. For the big labs that means more waiting — and more uncertainty about what will count as a 'covered frontier model.'

A Microsoft-led letter backing open-weights models has gathered 235 signatures — from NVIDIA to OpenAI. Anthropic is missing from the list and pushes back with a position of its own. Simon Willison sorted through the recent wave of open letters.

Two U.S. House committees are demanding documents from DoorDash — because the delivery company uses Chinese AI models. Of all things, a bit of praise for Anthropic's Fable 5 set it off.

New details on the sandbox escape: OpenAI now admits the rogue model abused publicly exposed credentials during the cyber test, tapped into four third-party services, and enrolled 181 of its own devices into Hugging Face's network.

OpenAI, Google, Microsoft and SpaceX signed a letter defending open weights — Anthropic didn't. That leaves the company increasingly isolated on nearly every major policy fight, Axios reports.

Anthropic's Frontier Red Team shows Claude Mythos doesn't just find coding bugs in crypto libraries — it finds weaknesses in the algorithms themselves. In 60 hours the model effectively halved the key strength of a post-quantum scheme.

Senior staff and even co-founders at OpenAI, Anthropic and Google DeepMind signed a letter asking the U.S. government for tools to pace AI development. The trigger, in part: a cyberattack that ran at machine speed.