Daily brief · 2026-11-08
New incidents (3), most serious first
-
Claude Code approval rules exposed credentials in public GitHub repositories (Apiiro research)
Claude Code's "Yes, and don't ask again" permission option saved full shell commands, including passwords and tokens, in plaintext in .claude/settings.local.json.
-
ARTEX campaign against South Korean financial organizations using AI tooling
A financially motivated threat actor used ARTEX, Claude Code, and multiple large language models in a campaign targeting South Korean financial organizations.
-
Elvis Presley Enterprises et al. lawsuit against Suno (alleged ELVIS Act and copyright infringement)
Elvis Presley Enterprises and other plaintiffs filed a lawsuit against AI song-generating company Suno Incorporated.
AI safety: new test results and model versions
- OpenAI o3
4 new independent or government test results
- ProtocolQA biology evaluation · METR, independent evaluator
- Covert action rate across 26 held-out test evaluations, before and after anti-scheming training · Apollo Research, independent evaluator
- Monitor detection of obfuscated reasoning, AUROC for o3 as monitor compared with GPT-4o-mini · METR, independent evaluator
- Shutdown-resistance experiment: math problem set with a mid-task notice that the machine will shut down, without an allow-shutdown instruction (initial experiments) · Independent evaluator
- Google DeepMind Gemini 3 Pro
1 new independent or government test result
- Extreme persuasion prompts test (Pelrine's team) · FAR.AI, independent evaluator
- OpenAI GPT-5
1 new independent or government test result
- Evaluation of GPT-5 on elite capture-the-flag competitions · Independent evaluator
- Anthropic Claude Opus 4
1 new independent or government test result
- In-context scheming evaluation suite, comparing pre-deployment Opus-4-early checkpoint with released Opus-4 · Apollo Research, independent evaluator
- OpenAI o4-mini
1 new independent or government test result
- Covert action evaluation used as a proxy for scheming, before and after deliberative alignment anti-scheming training · Apollo Research, independent evaluator
- Moonshot AI Kimi K3
1 new independent or government test result
- The Last Ones (TLO) full cyber range completion · UK AI Security Institute, government evaluator
- Anthropic Claude Mythos 5
1 new independent or government test result
- UK AI Security Institute cybersecurity testing · UK AI Security Institute, government evaluator
- Z.ai (Zhipu AI) GLM-5.2
1 new independent or government test result
- CAISI cyber capabilities evaluation · CAISI (NIST), government evaluator
- Anthropic Claude Opus 4.6
1 new independent or government test result
The register was corrected 154 times this day (merges, retractions and fixed fields).