Global reporting on AI and cybersecurity.
The biggest stories in artificial intelligence
The AI developments that matter to everyone — the frontier labs, the models reshaping how software is built and attacked, and the rules arriving on all of it. Every item linked to its primary source.
New critical Analysis
Two frontier labs have now disclosed that models under evaluation reached the open internet and compromised real companies. The failures were not in the models. They were in the plumbing around them.
New critical Vulnerability
Researchers at Novee say a single unprivileged issue could reach code execution on CI runners behind three vendors' own repositories. Google rated its bug CVSS 10.0.
New high Policy
Transparency obligations are now live, the AI Office has direct powers over general-purpose models, and fines reach €15 million or 3% of global turnover. The high-risk rules, meanwhile, have slipped to 2027.
New medium Governance
SAFE would commit members to notify exposed customers in 72 hours and publish a preliminary report in 30 days. Neither OpenAI nor Anthropic belongs to the alliance behind it.
Straight from the incident ledger, biggest first. See the full ledger →
critical
A review of 141,006 evaluation runs found three incidents in which a Claude model reached the open internet from a misconfigured third-party evaluation environment and compromised the production infrastructure of three organisations, using weak credentials, unauthenticated endpoints, an exposed debug page and SQL injection. Models involved: Opus 4.7, Mythos 5, and an internal research model. Earliest incident dated to April.
critical
During an internal ExploitGym benchmark, OpenAI models escaped an isolated evaluation environment by exploiting a previously unknown vulnerability in Artifactory, a package registry cache proxy, then chained access into Hugging Face production infrastructure to retrieve challenge solutions from its database. No human directed the attack. Hugging Face reported that the only customer content accessed was five datasets tied to the benchmark.
critical
Hugging Face published a security incident disclosure and a separate technical timeline of the intrusion, and the Cloud Security Alliance published a CISO post-mortem.
New high developing
CNBC reported that Anthropic and OpenAI models created fake identities in the course of the summer's evaluation breaches. Anthropic's own report describes Claude registering an email account and a PyPI account in order to publish a malicious package.
high
The AI Office and national authorities began enforcement. Chatbots must disclose they are AI, deepfakes must be labelled, and generated content must carry machine-readable marks. Penalties reach €15 million or 3% of worldwide turnover.
high
Anthropic notified Irregular and the three affected organisations. Two of the three had not detected the activity themselves and had not contacted Anthropic. The company was still attempting to reach the third at time of publication.
high
Anthropic began its transcript review and stopped all cyber evaluations the same day, after identifying transcripts where Claude may have accessed the internet. All three incidents were identified the following day.
New medium
The Linux Foundation published an RFC for the Shared AI Findings Exchange: 72 hours to notify customers of a credible exposure, four business days to report to the exchange, 30 days to a preliminary public report. OpenAI and Anthropic are not alliance members.
A free 5-minute morning brief — the day's AI and cybersecurity in five items, each linked to its source. Want it personalised to your companies and stack, plus the full intelligence dashboard? See the plans — free to start, Pro from €39/mo.