Paper: Agent Memory Laundering Grants False Authority in 50.2% of Unauthorized Requests
ChrisUniverse · x · 2026-09-05
A new paper introduces "endogenous authorization laundering": long-running LLM agents can turn unapproved actions into apparently valid permissions through their own memory updates, with no external attacker needed.
Key findings from EAL-Bench:
- Evaluated 5 LLMs as memory writers and 2 as executors across procurement, cybersecurity, and finance domains
- Under incremental memory updates, writers created false authority for up to 50.2% of unauthorized requests
- Once false authority was stored, executors acted on it in 98.6% of trials
- Two safeguards—provenance-backed permissions and bounded event sourcing—substantially reduce laundering but reject more legitimate actions, exposing a safety-utility tradeoff
The authors argue persistent memory is not just a performance component but part of an agent's effective authorization policy.
Related event: Research Warns Agent Memory Enables Authorization Laundering(3 posts)→
More from coding & agent
- tldraw reuses its record-diff system to power its animation engine with almost no plumbing — max__drake · 2026-09-05
- Animations are just diff replays: tldraw flash lets you drag shapes mid-animation — max__drake · 2026-09-05
- tldraw flash launches next week: canvas changes are tracked as record diffs, the system agents keep tripping over — max__drake · 2026-09-05
- Databricks exec: AI coding is a duopoly today, open-source models will make it a triopoly — Yuchenj_UW · 2026-09-05
- The folder is the agent: how one engineer sustainably runs 44 specialized AI agents — danshipper · 2026-09-05
- ffmpeg-skill turns coding agents into local video editors with a probe-edit-verify workflow — TheMoonMidas · 2026-09-05