Encrypted Skills Can Become New Security Exploits If AI Agents Know How to Decipher

Evidence record

October 13 arXiv paper metadata/abstract corroborated from original bibliographic identifiers and independently mirrored abstract; limited direct full-text access.

Read the original source ↗

Claim-level citations (3)

• supportsEncoded instructions in skill files can be recovered and promoted into agent actions: The authors report peak attack success of 47.5% for a Caesar-style variant and 40.3% for a layered variant in their primary shortened-skill condition.

  Abstract; stated primary shortened-skill findings

• supportsEncoded instructions in skill files can be recovered and promoted into agent actions: The controlled study embeds encoded task instructions into untrusted skill files and tests Caesar and layered variants across eight language models and 139 injection-task pairs.

  Abstract; reported experimental design

• supportsEncoded instructions in skill files can be recovered and promoted into agent actions: Instruction recovery and malicious-action success are separate outcomes; successful decoding does not by itself mean the attack succeeded.

  Abstract; distinction between recovery and attack success