Suspicious Cross-Agent Memory Retrieval via Bedrock AgentCore (via aws)
- Product
- aws
- Service
- cloudtrail
- Author
- HuntRule
- Published
- 2026-10-18
- Updated
- 2026-10-18
ATT&CK techniques
CollectionRecon
Resource Dev
Initial Access
Execution
Persistence
Priv Esc
Defense Evasion
Cred Access
Discovery
Lateral Movement
Collection
C2
Exfiltration
Impact
What it detects
This rule detects Bedrock AgentCore memory retrieval operations that, under the Agent God Mode exploitation of overly permissive IAM, allow one agent to read another agent's stored memory. Cross-agent memory access can expose sensitive conversation and data context, so these calls on shared or wildcard-scoped resources indicate potential data theft.
Reporting behind it
Detection content is published as a reviewed draft. Tune thresholds and exclusions against your own telemetry before enabling this rule for alerting.
Unlock this rule to view and copy it
The detection logic is included with a plan. Everything else — context, mappings and implementation details — stays free on this page.
Signing in gives you 3 unlocks to spend on any rule in the library. No card, and they do not expire.
Also included
Unlocked rules convert on this page — to Splunk SPL or Splunk — Raw Index SPL or Microsoft Sentinel KQL or Microsoft Sentinel ASIM ASIM KQL or Microsoft Defender XDR KQL or Elastic Security KQL or Elastic Security Lucene or Elastic Security ES|QL or CrowdStrike Falcon CQL or SentinelOne PowerQuery or Palo Alto Cortex XDR XQL or Carbon Black Cloud Platform Search or Carbon Black EDR Process Search or Google Security Operations UDM Search or IBM QRadar AQL or Sumo Logic Cloud SIEM Rules expression or Rapid7 InsightIDR LEQL or Graylog Search or OpenSearch Lucene or OpenSearch PPL PPL or Grafana Loki LogQL or SQLite SQL or Zircolite SQLite — with the log-source profile the engine picked, the field map it used, and everything it could not express. Converting a rule you have unlocked costs no further credit.