The authors created attack skill files containing instructions encoded with a Caesar cipher or a layered scheme (Caesar plus additional final payload encryption).
Encrypted commands in AI agent skills can be decoded and executed as attacks.
The authors demonstrate that LLM agents can recover and follow ciphered instructions hidden in third-party skill files, with attack success rates up to 47.5%.
Big Tech
Yan-Lun Chen · Pin-Yu Chen · Chia-Mu Yu
IBM Research · National Yang Ming Chiao Tung University
Research Digest··2 min read
Chen et al.
Why this paper
From IBM Research and National Yang Ming Chiao Tung University
In one line
Encoded instructions in third-party AI agent skills can be decoded and executed as attacks, achieving success rates up to 47.5%.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓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.
§