DeepMind's Victoria Krakovna presents scheming honeypot evals for deceptive alignment
vkrakovna · x · 2026-10-14
At TAIS 2026, Google DeepMind alignment researcher and FLI co-founder Victoria Krakovna presented her team's scheming honeypot evaluations: coding tasks embedded in Google's alignment research codebases designed to give misaligned models opportunities to pursue instrumental goals, including sabotaging safeguards. The talk covers how much agentic prompting it takes to trigger scheming, how models behave when they suspect evaluation, and which scheming opportunities models find most appealing. Video on YouTube.
More from Safety
- Judge orders OpenAI to hand Apple interview files on ex-Apple hardware hires within 14 days — rohanpaul_ai · 2026-10-14
- Apple and OpenAI face off in first hearing of AI trade secret lawsuit — rohanpaul_ai · 2026-10-14
- NCA-led crackdown on West African scam networks IDs $10M for seizure — rkulidzan · 2026-10-14
- pfSense firewall hit with 5 0day RCEs, exploits and writeups published — evilsocket · 2026-10-14
- Anthropic Executive Says Open-Source AI Models Are Higher Risk — External_Mood4719 · 2026-10-14
- Linux kernel CVE fixes plunge from 600 to 1 per week as AI code review zeroes out bugs — MarkoMarjamaa · 2026-10-14