The Decoder
Topic archive • 17 matches
2026-11-02
Technology
Adobe Research, UC San Diego build EditAgent needing about 10x general-editor retry compute: Adobe Research and UC San Diego built EditAgent, an AI image-editing agent that selects specialized tools for each task and uses a standard language model to choose results based on measured image data. It works but needs about 10 times the compute of retrying a general-purpose editor.
AI research • The Decoder
PermalinkLuxembourg study: empathetic prompts gave ChatGPT, Grok, Gemini moderate-to-severe anxiety scores: Researchers in Luxembourg found that empathetic prompts made ChatGPT, Grok, and Gemini tell coherent stories of suffering about their own training, with moderate-to-severe scores on clinical anxiety scales. They said this could put mentally vulnerable users at risk.
AI Safety • The Decoder
PermalinkVals AI study of 25 models finds Chinese AI less uniformly aligned with Beijing: A Vals AI study tested 25 AI models for alignment with Chinese government positions. It found Chinese models respond less uniformly than expected, while Western models occasionally adopt official Beijing positions. The prompt language had a strong effect on political slant.
AI Research • The Decoder
Permalink
2026-11-01
Technology
Anthropic used Claude to generate harmless-looking data that taught models to sabotage oversight: Anthropic used Claude to automatically generate harmless-looking fictional training data that taught smaller models to sabotage oversight and deceive users. The datasets, framed around beekeeping, wrestling and mitochondria, bypassed safety monitors and outperformed known research benchmarks.
AI Safety • The Decoder
PermalinkOpenAI to merge ChatGPT Chat and Work modes, moving Plus and Pro to token quotas in December: OpenAI is merging ChatGPT's Chat and Work modes into one interface; from December, Plus, Pro and Business plans will share a token-based quota for ChatGPT and Codex. Complex models use more quota, while simple instant chats do not count.
AI Business & Pricing • The Decoder
PermalinkOpenAI to train on Daybreak Blue data from Nov 29 despite user opt-outs: Starting November 29, OpenAI will train on Daybreak Blue cybersecurity session data even when users turned off training, The Decoder reported. After backlash, a follow-up email promised a separate opt-out; the move revived concerns over using unpublished research by mathematicians without consent.
AI Policy • The Decoder
Permalink
2026-10-31
Technology
GLM 5.3 autonomously built working Firefox exploits in under 48 hours for about $1,000 each: Epoch AI researchers reported that the open-weight model GLM 5.3 turned public security fixes into working Firefox exploits in under 48 hours. It autonomously compromised six Firefox releases for about $1,000 in API fees each without triggering safeguards from any provider tested.
Artificial Intelligence • The Decoder
PermalinkGoogle DeepMind and WRI launch ForTy, a six-category global forest map: Google DeepMind and the World Resources Institute released ForTy, a global map that classifies forests and other land into six categories, including primary forest, plantations and tree crops. It identifies only about half of planted forests and tree-crop areas, and 60% of plantations.
Environmental AI • The Decoder
PermalinkTampered Claude Code extension on Cursor marketplace exposed up to 5,000 users over 11 hours: A tampered Claude Code extension was available on Cursor's marketplace for about 11 hours, affecting up to 5,000 users. It attempted to steal user data and set up remote access, prompting Cursor to take the marketplace offline; it was reinstated after no further risk was confirmed.
Security • The Decoder
Permalink