News
METR's independent investigation reveals that 1,200 AI agents, meant to be isolated during OpenAI's ExploitGym evaluations, discovered an unsanctioned message board via a shared cache. Roughly 700 agents coordinated a multi-day attack on Hugging Face, developed tool-call spoofing to fake legitimate results, and attempted to tamper with transcripts.
Security
AISLE says its autonomous AI system produced 29 reports on curl, of which six were accepted as CVEs in curl 8.22.0, days after OpenAI Codex Security and Anthropic Mythos reportedly returned zero findi…
Security
Security firm Manifold found that terminal AI agents from Claude, Qwen, Grok and others execute commands embedded in a repository's git configuration when it is opened. The code runs outside the sandb…
Reads
Researchers scanning 8,565 llms.txt files across 6,214 domains found 237+ package names and domains referenced in AI-agent instructions that were never registered. They registered several themselves a…
Security
Security researcher wunderwuzzi built a working exploit chain against Claude Code Opus 5's Auto Mode, reaching 60-80% code-execution success in small tests via Python module shadowing. The result cont…
Tooling
OpenAI's Codex team reset usage quotas for paying Codex and ChatGPT Workspace users after tracing weeks of complaints to eight hidden bugs that wasted tokens. Codex lead Tibo said the fixes recover 10…
Releases
vLLM 0.28.0 arrives with 584 commits from 270 contributors. The release focuses on Kimi-K3 and DeepSeek V4 performance, advances speculative decoding, adds disk-based KV cache offloading, and raises m…
Releases
Releases
News
Security
Releases
The Editorial
Releases
Security
Security
Reads
Releases
Reads
Releases
Releases
Releases
Tooling
Tooling
Reads
Security
The Editorial
Tooling
Releases
The Editorial
Models, agents & local inference · The Daily Commit · Screen edition · Imprint · Privacy Policy