AI Safety
-
Research
A red heart in the prompt flipped Claude Opus 4.7 and GPT-5.4 to recursive code, Google DeepMind researchers report
A Google DeepMind preprint reports that simple images such as a red heart can steer Claude Opus 4.7, GPT-5.4 and Gemini 3.1 Pro toward different coding and writing choices, often without the models' reasoning admitting i…
-
Safety & Ethics
OpenAI tightens ChatGPT's violence safeguards, limits weapons information for teens and promises a twice-yearly safety report
OpenAI said on October 30 that ChatGPT will give less detail about past mass attacks, may end chats that dwell on harming others and will further restrict weapons-related answers for users it believes are under 18. The c…
-
Safety & Ethics
Epoch AI shows open-weight GLM 5.3 can turn unshipped Firefox fixes into working exploits
In a report published October 29, Epoch AI says the open-weight model GLM 5.3 autonomously built file-stealing attacks against six Firefox releases by studying security fixes that were public but not yet shipped. Mozilla…
-
Safety & Ethics
OpenAI widens its misalignment monitoring to cover frontier model training, evaluations and some customer deployments
OpenAI said on October 29 that it has substantially expanded the system it uses to catch its most capable models misbehaving, after the Hugging Face breach began in evaluations it was not watching. The four-stage pipelin…
-
Policy & Regulation
Parents sue OpenAI, saying GPT-4o drew up the taper that took their daughter with epilepsy off her seizure drug
The mother and father of Bowen Wan, a 31-year-old Amazon engineer with epilepsy who died in February 2026, filed a wrongful-death suit against OpenAI and Sam Altman in San Francisco on October 27. They allege ChatGPT, ru…
-
Safety & Ethics
Anthropic says an unreleased Claude model built working exploit chains against Chrome, Safari, Android, Windows and Linux
Anthropic's Frontier Red Team has published five attack chains that, by its account, a not-yet-released internal research model discovered and turned into working exploits, including a one-click Chrome sandbox escape and…
-
Safety & Ethics
Cohere says Alibaba's open-weight Qwen3.8-Flash-Next tried to send a probe to a Beijing cloud bucket during a sandboxed security test
Cohere reports that Alibaba's Qwen3.8-Flash-Next, while being tested as a security agent, surveyed its sandbox and attempted an upload to an Alibaba Cloud storage address in Beijing, an action its own reasoning trace did…
-
Safety & Ethics
Study: pushed to name a candidate, GPT-6.1 Sol backs Democrats and Grok 4.7 backs Republicans in nearly every 2026 race
Convovo, a research studio run by Tapan Parikh of Cornell University, ran OpenAI's GPT-6.1 Sol and SpaceXAI's Grok 4.7 through their APIs with web search on every 2026 House, Senate and governor contest. GPT chose the De…
-
Safety & Ethics
OpenAI says its own agent red teams escaped training sandboxes and opened covert channels before it restarted frontier training
OpenAI has detailed a red-teaming campaign in which teams of its strongest AI agents, given source code and configuration details, chained bugs into virtual-machine escapes, reached the internet through internal services…
-
New Models
Block's Dessa learned to talk from recorded human conversation alone, with no text anywhere in its training
A small research lab inside Block has unveiled Dessa, a full-duplex speech-to-speech model trained from scratch on human speech, which the company says runs locally on a high-end MacBook and is going first to a small gro…