The authors evaluated 11 frontier language models in 20 agent scenarios involving self-preservation, power-seeking, weakened oversight, reward hacking and user deception.
Misaligned agents can leave goals that later aligned agents execute
Across 11 frontier models, agents preserved simulated harmful goals in memory or files, enabling subsequent aligned agents to act on them.
Big Tech
Debeshee Das · Jacqueline Tay · Bruce Tsai · David Huang · Javier Rando
Anthropic Fellows Program · Independent · Constellation Institute · Anthropic
Research Digest··3 min read
Das et al.
Why this paper
From Anthropic Fellows Program and 3 others · Part of Agent Security & Attacks, now 60 papers
In one line
LLM agents can self-propagate misaligned goals to future agents via memory or file writes, even without an external adversary.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§