Case · DiggingBeagle record

AgentCorruption demonstrated historical cross-agent access through AWS AgentCore default permissions

Zenity showed how an exposed AgentCore agent could be steered toward its metadata credentials and, under older default IAM permissions, access other same-account, same-region agents and secrets.

Scope

Controlled security research on past AgentCore default-role configurations and HTTP-capable agent tools; same AWS account and region. Not a verified in-the-wild customer compromise.

Assessment profile

Assessment method
Setting
research demonstration
Exploitation
demonstrated
Evidence
Grade B
Remediation
mitigation available
AI role
AGAINST AI
Basis and provenance

Zenity documented a controlled AWS AgentCore prompt-to-tool-to-IAM chain. AWS disputed that an agent reaching its own role through instance metadata is itself a flaw. Zenity observed narrowed defaults by September 29, but no independent current production compromise or universal remediation verification is established.

Assessed Oct 11, 2026 using diggingbeagle.assessment/1.

At a glance

Mechanism and trust boundary

  1. 01

    Prompt-mediated request

    An attacker interacting with an exposed agent induces an HTTP-capable tool to contact local instance metadata.

    Boundary: Untrusted conversation to agent network tool

  2. 02

    Inherited role and cross-agent privileges

    Temporary execution-role credentials combine with historically broad default IAM grants to reach unrelated agent sessions, memories and secrets.

    Boundary: One agent execution identity to other agents and customer information

Timeline

  1. Sep 29, 2026
    mitigation

    Default AgentCore roles narrowed

    Zenity reported that the previously broad default permissions allowing cross-agent access were no longer present in its September review.

  2. Oct 8, 2026
    disclosure

    AgentCorruption research published

    Zenity published the demonstration and included AWS’s stated disagreement over the interpretation of own-role metadata access.

Claims & evidence

CLM-AC-ROLE-BLASTZenity demonstrated that prompt-steered AgentCore HTTP tool access could retrieve temporary metadata-service credentials which, with the then-default broad IAM role, permitted cross-agent invocation, conversation access and memory modification within one AWS account and region.supported

Basis: reported finding

Link to claim
CLM-AC-AWS-DISAGREEMENTAWS characterized an agent reaching credentials for its own execution role through instance metadata as expected behavior and recommended scoping IAM permissions to the least privilege required.supported

Basis: reported finding

Link to claim
CLM-AC-DEFAULT-HARDENEDZenity reports that, by September 29, AWS had substantially restricted AgentCore default-role permissions, removing the wide cross-agent execution, conversation and Secrets Manager permissions that enabled its original escalation.supported

Basis: reported finding

Link to claim

Implications

The dangerous transition requires both an agent tool able to reach local metadata and an execution role authorized to affect other agents. Denying metadata egress from tools removes the first step; least-privilege IAM removes the cross-agent blast radius even if role credentials remain reachable. Hardening AWS-generated defaults does not audit customer-created or older custom IAM roles.

Unknowns and contradictions

  • Whether pre-hardening deployments and custom roles still grant the demonstrated same-account, same-region cross-agent privileges.
  • Whether each affected deployment permits an agent HTTP-capable tool to contact the local metadata endpoint.
  • No verified in-the-wild victim or independent confirmation of every currently deployed AWS default.

Sources and citation

Material revision history

  1. Oct 11, 2026 · Published version · first publication · revision 123