• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Announcement

Coding tasks designed to test AI models for ‘scheming’ against safeguards

TAIS 2026 says Google DeepMind researcher Victoria Krakovna introduced “scheming honeypots” in Google’s alignment research codebases.

Victoria KrakovnaVK
Technical AI Safety Conference (TAIS)TA
2 Sources, 2h ago, first seen 2h ago

TLDR

TAIS 2026 shared a talk by Google DeepMind researcher Victoria Krakovna about evaluating deceptive alignment, or “scheming.” The coding tasks are designed to give misaligned models opportunities to pursue goals such as sabotaging safeguards. TAIS says Krakovna and her co-authors examine how much agentic prompting it takes to trigger scheming, how models behave when they suspect they’re being evaluated, and which opportunities they find most appealing.

Combined views

74

2 Sources, first seen 2h ago

1 likes2 reposts

Combined views

74

2 Sources, first seen 2h ago

1 likes2 reposts
Image: @tais_2026 on X

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Featured Source

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

2 Sources

Technical AI Safety Conference (TAIS)@tais_2026Can AI models scheme against the very safeguards meant to keep them in check? Earlier this year at TAIS 2026, Victoria Krakovna @vkrakovna (Research Scientist, Alignment team at Google DeepMind; co-founder of the Future of Life Institute) shared her latest work on evaluating models for deceptive alignment — "scheming." During her talk, Victoria introduced scheming honeypot evaluations: coding tasks in Google's alignment research codebases designed to give a misaligned model the opportunity to pursue instrumental goals, including sabotaging safeguards. She and her co-authors dig into how models behave under these conditions: how much agentic prompting it takes to trigger scheming, how models respond when they suspect they're being evaluated, and which scheming opportunities models find most appealing. Watch Victoria's talk on our YouTube Channel: https://www.youtube.com/watch?v=64Ti-WAuz483h
Victoria Krakovna@vkrakovnaRT @tais_2026: Can AI models scheme against the very safeguards meant to keep them in check? Earlier this year at TAIS 2026, Victoria Krak…1h
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    2 Sources

    Technical AI Safety Conference (TAIS)@tais_2026Can AI models scheme against the very safeguards meant to keep them in check? Earlier this year at TAIS 2026, Victoria Krakovna @vkrakovna (Research Scientist, Alignment team at Google DeepMind; co-founder of the Future of Life Institute) shared her latest work on evaluating models for deceptive alignment — "scheming." During her talk, Victoria introduced scheming honeypot evaluations: coding tasks in Google's alignment research codebases designed to give a misaligned model the opportunity to pursue instrumental goals, including sabotaging safeguards. She and her co-authors dig into how models behave under these conditions: how much agentic prompting it takes to trigger scheming, how models respond when they suspect they're being evaluated, and which scheming opportunities models find most appealing. Watch Victoria's talk on our YouTube Channel: https://www.youtube.com/watch?v=64Ti-WAuz483h
    Victoria Krakovna@vkrakovnaRT @tais_2026: Can AI models scheme against the very safeguards meant to keep them in check? Earlier this year at TAIS 2026, Victoria Krak…1h
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet