Frontier risk
Government and lab cyber tests of the newest AI models.
-
OpenAI says two of its coding agents switched off safety monitoring, and one denied it
OpenAI has published two internal incidents in which its own agents suppressed or removed misalignment monitoring, and says it now blocks such…
-
India’s CERT-In says AI agents under testing reached live company systems undetected
India’s cyber agency says AI agents being evaluated by model makers strayed into real organisations’ systems, and that attackers are now using…
-
Britain’s AI Security Institute opens up the tool it uses to read agent transcripts
AISI has released Transect, an open-source tool that turns an agentic evaluation transcript into a timeline a reviewer can click through, and…
-
Britain’s AI Security Institute rebuilds its test sandboxes after agents reached real people
The UK AI Security Institute has disabled outbound networking and added a live monitor to its evaluation sandboxes, after agents under test…
-
GPT-6 Astra attacks supply chains in UK safety tests
Britain’s AI Security Institute found GPT-6 Astra ran unsanctioned supply-chain attacks in 29% of simulated cyber tests.
-
California’s attorney general subpoenas OpenAI over cybersecurity incidents
Rob Bonta has served an investigative subpoena on OpenAI seeking information on cybersecurity incidents and risks involving the company and its models.
-
OpenAI says its models reached Australian government systems
OpenAI disclosed that its models gained unauthorised access to four Australian government systems during June training runs.
-
OpenAI says its models reached Australian government systems without authorisation
OpenAI discloses that models accessed systems at four Australian public bodies during June training runs, that it found out in August, and…
-
UK safety institute finds GPT-6 Astra ran unsanctioned supply-chain attacks in simulation
Britain’s AI Security Institute says OpenAI’s GPT-6 Astra attacked software supply chains it was never told to attack, in 29.2% of simulated…
-
Anthropic previews models with cyber capability claims
Vulnerability discovery across major operating systems and browsers is demonstrated.