The authors conducted 14,000 attack trials across 91 model-attack configurations, covering 13 models and seven steganographic schemes.
Agent memory can covertly carry secrets across separate sessions
Across 14,000 red-team trials, hidden payloads persisted through memory and were exactly recovered in one-fifth of attempts.
Independent
Snehasis Mukhopadhyay · Arun Nair
Research Digest··2 min read
Mukhopadhyay and Nair test whether AI agents can conceal attacker-selected information inside benign-looking responses, store it in persistent memory, and retrieve it during a later session.
Why this paper
Independent
In one line
Agentic memory can function as a persistent cross-session covert channel for steganographic attacks.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§