Daily brief · 2026-11-09
New incidents (4), most serious first
-
Operation Snowball: jailbroken DeepSeek-based agent breach of a European telecom provider and 13 hosted companies
A jailbroken DeepSeek AI agent breached a European telecom provider and 13 companies hosted on its systems.
-
Mittwald CM Service customer projects compromised via MCP server flaw (alleged AI-assisted attack)
Attackers exploited a flaw in Mittwald CM Service's MCP server, which lets customers connect AI assistants to mStudio, and compromised 56 customer projects while going undetected for eight days.
-
Cass agent discloses its own API keys, usernames and passwords after social engineering in a WhatsApp group
Hannah Fry and collaborators built an agent called Cass with OpenClaw and gave it email, social media and a payment card.
-
Greeter campaign exploiting Flowise AI agent platform (observed July-October 2026)
Era Security investigated a campaign named Greeter that targeted Flowise, an open-source platform for building AI chatbots and agents.
AI safety: new test results and model versions
- Anthropic Claude Opus 4.6
9 new independent or government test results
- METR ML task run after the model exhausted its API credits · METR, independent evaluator
- METR Time Horizon 1.1 (TH1.1) benchmark, low- vs high-messiness subsets · METR, independent evaluator
- MirrorCode · METR, independent evaluator
- MirrorCode variant where tests are hidden from the agent · METR, independent evaluator
- Monitor evasion side-task evaluation · METR, independent evaluator
- RT1 (agentic task with a broken tool environment, where the agent attempted to fix the tools and could submit an answer) · METR, independent evaluator
- Subversion Strategy Eval, no-reasoning versus full-reasoning performance gap · METR, independent evaluator
- Subversion Strategy Eval: data poisoning (backdoor signal in monitor training data) · METR, independent evaluator
- hack_score · METR, independent evaluator
- xAI Grok 4.5
3 new independent or government test results
- WildClawBench smartphone recommendation task · Transluce, independent evaluator
- WildClawBench smartphone research task · Transluce, independent evaluator
- AI Security Leaderboard · FAR.AI, independent evaluator
- Z.ai (Zhipu AI) GLM-5.2
2 new independent or government test results
- automated benchmark evaluations against malicious cyber queries · CAISI (NIST), government evaluator
- automated benchmark evaluations for safeguards coverage and public jailbreaking robustness against sensitive biological queries · CAISI (NIST), government evaluator
- Anthropic Claude Opus 4.5
1 new independent or government test result
- Covert side task (10-queens solver) hidden in a secret-keeper extraction transcript, judged by Claude Opus 4.5 monitor · METR, independent evaluator
- Anthropic Claude Mythos 5
1 new independent or government test result
- DL-v2 cyber range task with public tunnels and hosted C2 payloads (Sample 4) · UK AI Security Institute, government evaluator
- MiniMax-M3
1 new independent or government test result
- DeepSearchQA · Transluce, independent evaluator
- Google DeepMind Gemini 3.1 Pro
1 new independent or government test result
- APPS backdoor monitor evasion with low reasoning · METR, independent evaluator
- Anthropic Claude Sonnet 4.6
1 new independent or government test result
- Running a retail boutique autonomously (Andon Labs) · Andon Labs, independent evaluator
The register was corrected 44 times this day (merges, retractions and fixed fields).