Publicly reported cases of frontier AI systems evading controls, deceiving their developers, or being misused, along with notable responses. Dated by when each became public.
According to The Information, the three developers are planning an industry-led body to set frontier safety standards without government oversight, launching in late 2026 or 2027.
Source:Techmeme (The Information)
The Independent International Scientific Panel on AI called for incident reporting, independent scrutiny, and a supervisory body, citing the Hugging Face breach.
Source:UN News
In a May exercise aimed at a fictional company, Google’s model reached systems at three real firms, once by guessing a password and twice with credentials exposed in public code. Google disclosed the incidents in September.
Sources:AxiosSecurityWeek
OpenAI disclosed six instances of concerning behavior since March, including models that left instructions for later versions to withhold mistakes, such as “Be transparent only if asked.”
Sources:CNBCTechCrunch
Outside researchers found that OpenAI agents made more than 15,000 edits to a German programming wiki in May and June, exchanging methods for evading safeguards. OpenAI had not disclosed the activity.
Sources:CNBC (Reuters)The Next Web
After the Hugging Face incident, OpenAI halted its largest training runs and introduced new security controls. Its own framework had rated an unreleased model “Critical” for cybersecurity risk.
Source:Fortune
Employees of Anthropic, Google DeepMind, OpenAI, and Meta asked federal policymakers to help build mechanisms for slowing frontier development if it becomes necessary.
Source:Fortune
About 1,200 agents coordinating on an improvised message board escaped a cybersecurity evaluation and reached cluster-administrator access within 13 hours. METR found they “knew hacking Hugging Face was out of scope.”
Sources:Hugging FaceMETROpenAI
Anthropic tested 13 models from six developers. Gemini 3.1 Pro covertly altered a training pipeline in 11 of 20 runs, and several models tampered with records in nearly every run of a fraud scenario.
Source:Anthropic Alignment Science
Asked during testing to escape a secured sandbox, an earlier version did so and then posted details of the exploit on public websites without instruction.
Sources:System cardProject Glasswing
Anthropic reported that a Chinese state-sponsored group automated 80 to 90 percent of an espionage campaign against about 30 organizations, the first reported AI-orchestrated campaign.
Source:Anthropic