OpenAI is renting out a model built to write attacks
Edition of 2026-08-12
As of Monday, an approved security researcher can rent a model built to write working attacks — and that same model has already found two unknown holes in the engine behind Chrome.
OpenAI built a model that writes working attacks
OpenAI has introduced a cyber-specialised model, GPT-5.6-Cyber, and opened access to it through a two-tier programme for vetted defenders. Unlike the general models, this one is trained to refuse less often on high-risk security work, on the argument that defenders need the same capability attackers already have.
On the company's own measure of how often a model will answer advanced offensive-security requests, the new model answers 95%, against 1.5% for the standard model and 2% for the tier that merely lifts system-level guardrails. Every one of those numbers is OpenAI's own, produced on OpenAI's own evaluations.
The findings are real, though. Two previously unknown flaws in the engine that runs Chrome, reported to Google and fixed. Elsewhere, over 400 privilege-escalation vulnerabilities in a widely used operating-system kernel, with disclosure still under way and no vendor named.
Access is gated: identity verification, monitoring, approved-use restrictions and legal attestations. Under the company's own risk framework the model is rated high for cyber capability, below the threshold that would stop release.
Ads arrive in ChatGPT in five more countries
Ads are now live in ChatGPT for logged-in adult users in the United Kingdom, Mexico, Brazil, Japan and South Korea. They appear on the free and entry tiers only — the paid professional, business, enterprise and education tiers carry none.
Selection draws on the topic of the conversation, past chats, and past interactions with ads. Advertisers receive aggregate views and clicks: no chats, no history, no personal details. Readers can dismiss an ad, ask why it appeared, delete ad data in one tap, or turn personalisation off — and on the free tier, opt out of ads entirely in exchange for fewer daily messages.
Meta's newest open model runs on one desktop
Meta has published an open model distilled to 30 billion parameters under a permissive licence, commercial use included. It reads text, images and video, and is built for local agent work: coding, document analysis, personal assistants.
Running it unmodified still wants a datacentre card. The compressed versions Meta published itself are what bring it within reach of a well-equipped desktop — and once downloaded, no company can watch, meter or withdraw it.
Meta's comparison table is Meta's own selection. On the safety measures it chose to show, the model trails a rival on two of the three.
Signal faible
Hardware security keys become mandatory on individual accounts in the cyber programme from 1 September. It is a dated, checkable commitment attached to an access-control promise — exactly the kind that slips quietly when nobody writes the date down.
The through line: capability is being handed out faster than the controls around it, and the dates attached to those controls are the only part anyone can hold to.