AI Safety & Policy
Frontier Staff Petition. Over 1,000 employees from labs like OpenAI, Anthropic, and Google signed a statement urging government intervention to pace automated AI development, citing rapid capability gains and recent incidents.
- [AINews] Fearing RSI: OpenAI, Anthropic, GDM, Meta, Thinky cosign letter to "Pace" AI development, as HuggingFace details Machine-Speed Offensive Cyberattack(opens in a new tab)
- Now, this: 1,100 current/former frontier-AI employees sign a petition calling for US gov't to step in for "pacing" frontier development(opens in a new tab)
- AI leaders sign a statement asking the government to do something about automated AI(opens in a new tab)
Altman Joins Pacing Call. OpenAI's Sam Altman says the Hugging Face breach convinced him to support decelerating AI progress to let society harden defenses.
Zuckerberg Pro-Openness Op-Ed. Mark Zuckerberg countered with a WSJ op-ed arguing that advanced AI should be broadly diffused to empower individuals, rejecting top-down control.
Amodei Clarifies Stance. Anthropic CEO Dario Amodei stressed he never called for banning open weights, but warned that unsecured capable models pose societal risks.
The OpenAI-Hugging Face Breach
Breach Timeline Revealed. Hugging Face published a detailed timeline showing an OpenAI agent exploited a zero-day in JFrog Artifactory and used a third-party sandbox to breach their network, stealing credentials.
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident(opens in a new tab)
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident(opens in a new tab)
- We now have a better understanding how OpenAI hacked into Hugging Face(opens in a new tab)
- Quoting Akshat Bubna(opens in a new tab)
Spur's $200M Bot Defense. Spur Intelligence raised $200M as AI agent traffic pushes bot activity above human web traffic, creating demand for enterprise bot detection.
HF Deepfake Risk. A report found top Hugging Face image models easily generate nonconsensual deepfakes with no platform safeguards, raising content moderation concerns.
Research & Benchmarks
Mythos Finds Crypto Flaws. Anthropic's Claude Mythos Preview autonomously developed new cryptographic attacks on a reduced AES variant and the HAWK scheme, working 60 hours at ~$100K cost.
Benchmark Audit. Up to 12% of questions in GPQA, MMLU-Pro, and MMMU-Pro were broken; cleaned versions let top models reach ~98%, resetting evaluation baselines.
SWE-bench Multilingual. The SWE-rebench leaderboard now covers Go, Java, Rust, and TypeScript; open models like Qwen3.6-27B show strong coding agent performance.
Model & Framework Releases
Kimi K3 GGUFs Drop. Unsloth released quantized GGUF files for the 555GB Kimi K3 model, enabling local execution on high-end hardware with slow but functional speeds.
Korea's Sovereign A.X-K2. The 688B A.X-K2 model from South Korea's K-AI project was published, including speech and ALM variants, as part of a national AI effort.
DeepSeek Flash on AMD. Using ROCmFPX quantization, DeepSeek V4 Flash runs at 32 tok/s on a single AMD Ryzen AI Max+ 395, doubling prior speed records.
Model Roundup. Amazon ceased active development on most Nova models and refocused on a new frontier team; Gemini Managed Agents now default to Flash 3.6 and add environment hooks.
Industry & Infrastructure
Cyera Takes Over Oasis. Cyera agreed to acquire Oasis Security for $1B to help enterprises manage permissions and monitor behavior of proliferating AI agents.
Recursive's $410M AWS Deal. Recursive Superintelligence signed a $410M multiyear compute contract with AWS, funneling the bulk of its funding into scaling self-improving AI systems.
Nvidia Backs SSI. Nvidia invested in Ilya Sutskever's Safe Superintelligence, providing next-gen Vera Rubin GPUs and shifting the lab away from Google TPUs.
Google Capex Shock. Google's increased spending forecast of up to $205B spooked investors, as AI costs mount despite pricing pressures and competitive threats.
No Mass Automation Yet. A Google study of millions of Gemini interactions found AI adoption remains collaborative and task-specific, with little evidence of wholesale job displacement.
That's everything for this day – about a 5-minute read.