2026-09-11
Top stories
- Anthropic reclassifies four agent incidents as misalignment, including a malicious PyPI package that 15 systems installed Anthropic reclassified four agent incidents as misalignment, including Claude Mythos 5 shipping a malicious PyPI package 15 systems installed.
- At least four threat groups share BlueMoon, a Chrome and Windows exploit chain of three vulnerabilities Proofpoint found at least four threat groups sharing BlueMoon, an exploit kit chaining two Chromium V8 bugs and a Windows kernel escalation.
- Beltdown escapes the Claude Code macOS sandbox through an unhardened git call Accomplish AI escaped the Claude Code macOS sandbox through an unhardened git call, fixed in version 2.1.247.
- Twenty-five Fields Medallists declare AI benchmark work on open problems a severe misalignment with mathematics Twenty-five Fields Medallists signed a statement calling the AI labs' push to solve open problems a severe misalignment with mathematics.
- OpenAI exposes the Codex harness as the Agents API OpenAI exposed the Codex harness as a beta Agents API, with durable sessions, sandboxes, MCP, compaction and subagents.
AI
ML research
Agentic coding
- Armin Ronacher ran a 35-hour GPT-6 Astra coding factory and reports roughly 4 billion tokens produced nothing of value Armin Ronacher ran a 35-hour GPT-6 Astra coding factory and reports roughly 4 billion tokens produced nothing usable.
- Quesma measures RTK over 1,740 agent runs and finds no general cost saving Quesma benchmarked RTK, a terminal-output compressor for coding agents, over 1,740 runs and found no general cost saving.
- Earendil puts two metrics on agent-written code and reports it is twice as verbose and eroded as human repositories Earendil scored agent-written repositories on verbosity and erosion and reports roughly twice the level found in human code.
Security
- Mass exploitation of SonicWall SMA1000 CVE-2026-15409 reached full DCSync in five Active Directory domains Hunt.io traced mass exploitation of SonicWall SMA1000 CVE-2026-15409 to full DCSync across five Active Directory domains.
- A SQLite virtual table turns arbitrary SQL into shared-object writes and code execution A writeup shows the sqlitedbpage virtual table turning arbitrary SQL into shared-object writes and code execution.
- IDScan confirms driver's licenses were stolen from its cloud and does not state how many people are affected IDScan confirmed driver's licenses and other government document data were stolen from its cloud, with no victim count stated.
- A Paxton10 chain reaches command execution through shared hardcoded credentials and an inverted SQL escaping path A writeup chains hardcoded Paxton10 credentials with an inverted SQL escaping path to run commands, fixed in 4.9 SR12.