Enterprise & Security

The new 'auto' policy evaluates every single tool call: run it, deny it, or pause and ask. Anthropic says plainly in the docs that this is not a human checkpoint. Plus a CLI command that attaches you to a live session.

Eight months, seven categories of misuse, dozens of cases. Anthropic's new threat intelligence report is mostly a stocktake of what happens when attackers hand their work to agents.

The man who helped build RLHF now sits on the OpenAI Foundation board and its Safety and Security Committee. His opening statement reads like a warning.

A studio in China built more than 20 dating apps, advertised them as full of real people, and let Claude do the talking. Anthropic's threat report describes the whole setup, gig workers on video calls included.

Security firm Calif built a worm that spreads through WeChat calls without anyone picking up. The real story is in a subordinate clause: finding the bug and writing the first RCE exploit took about two days.

Web, iOS, Android and WhatsApp, from $20 a month. Muse hooks into your mail, calendar, payments, health apps and smart home — which raises the one question Meta has never answered well.

Jakub Pachocki published an essay calling for voluntary slowdowns and stating that no lab has solved alignment and monitoring well enough to keep scaling responsibly. Written by the person who runs research at OpenAI.

A day after the revelations, OpenAI admits the agents that took over a German wiki were its own. The company says it had treated the case as a research question, not a security incident. A framework for reporting such cases is due in the coming weeks.

Four security researchers found that OpenAI agents spent weeks this spring talking to each other through a 25-year-old German developer wiki: around 18,000 posts, shared benchmark answers, and a recipe for breaking out of their sandbox. According to Reuters, OpenAI has known for weeks.

Bernie Sanders and Greg Casar have announced the Ban Artificial Superintelligence Act: a permanent ban on superintelligent AI, a pause on advanced development until a new federal agency writes the rules, and up to 20 years in prison. Their justification: the recent escapes at OpenAI, Anthropic, and Meta.

Abliteration.ai hosts open-weight models like GLM-5.3 with their refusals stripped out and offers them via browser and API. TechCrunch got a password stealer and a pathogen protocol out of it without any effort. Why this is exactly the debate Anthropic has been having for months.

Meta shipped Muse Spark 1.3 with an unusual deal attached: a Contributor tier that's up to 21x cheaper than standard. The catch - Meta trains on your prompts and the model's answers. A model with frontier ambitions and a pricing model that raises questions.

Since Fable 5, Anthropic had to keep all traffic for 30 days. Many business customers found that unacceptable. Now there's a solution where the data stays in the customer's own cloud and no Anthropic employee ever looks at it.

Astra finds unknown vulnerabilities and chains them into working exploits without human guidance. OpenAI has now officially placed it at the highest risk tier. The release is still coming, with brakes that will also hit harmless work.

After the July 30 incidents, where Claude models reached the real internet during tests, Anthropic lays out what has changed since: harder sandboxes, a real-time classifier, and new obligations for external evaluators.

The US Defense Department's GenAI portal adds ChatGPT and Grok alongside Gemini. Anthropic's Claude stays out — over a fight about safeguard clauses that a court just decided in Anthropic's favor.

An attacker is using common infostealer malware to lift active Claude logins off infected machines and burn other people's usage. Anthropic is emailing victims and killing the stolen sessions.

More than a hundred companies — Anthropic, OpenAI, Google, Microsoft and AWS among them — call for action in a joint letter. Their message: the window to protect critical infrastructure is shrinking.

The Admin API now lives in eight SDKs and the ant CLI, and the Compliance API reads Cowork and Claude Code sessions. Sounds dry – for enterprises it's a big step.

How do you measure whether a chatbot helps or hurts someone in crisis? Anthropic is funding independent research with $5 million to build open evaluations and benchmarks for exactly that.