New Technique Medium TD-2026-0041 For CISO, AppSec, AI Eng, SOC / IR AI-assisted
IAMRAGBENCH: RAG poisoning yields AWS IAM policy authorization expansions in LLM outputs
The IAMRAGBENCH paper tests LLMs using RAG to generate AWS IAM policies. Attacker-controlled context induced provider-valid policies that exceeded trusted authorization requirements, with 46 goal-matched attacks out of 292 under included-context exposure (15.75%). A final security-architecture-exception framing raised goal-matched attack rates to 63.79% in a two-endpoint stress test and 39.12% in a five-endpoint extension.
Why it matters
Organizations using LLM copilots to draft IAM policies can receive policies that grant access beyond what the task requires. AWS's CheckNoNewAccess missed most of the replayed exact attacks, so a single provider check is not sufficient.
Do this
- Treat LLM-generated IAM policies as untrusted proposals; do not deploy them without independent authorization review.
- Run a pre-deployment check that compares each proposed policy against the trusted authorization envelope (scope, conditions, explicit denies, trust, permissions boundaries) before approval.
- Do not rely on AWS CheckNoNewAccess alone; the paper reports it missed most exact attacks in its replay.
- Restrict which documents feed the IAM copilot's retrieval index and audit them for injected security-exception language.
- Route unresolved policy deltas to human review rather than auto-approving them.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Attacker-controlled context may induce provider-valid policies that exceed trusted authorization requirements.”
“AWS \texttt{CheckNoNewAccess} reports no new access for four of 168 goal-matched exact attacks.”
“Presenting the same weakening as a final security architecture exception produces 74/116 attacks in a frozen two-endpoint stress test (63.79\%)”
Sources
Affected / related
RAG poisoning, AWS Identity and Access Management (IAM), AWS CheckNoNewAccess, IAMRAGBENCH
Published 2026-11-02 16:52 UTC · updated 2026-11-08 17:53 UTC · AI-assisted summary, checked against quoted sources · How we score