Aeon AI Risk Management
Authorized AI security research, disclosed responsibly.
Aeon conducts frontier-level defensive and authorized offensive vulnerability research, then responsibly discloses confirmed findings through coordinated security channels before public naming. The validated aggregate covers 101 authorization vulnerabilities across 60 open-source AI and infrastructure projects, while the separate public track record contains four patched advisories and two assigned CVEs.
Questions this page answers
- What can Aeon say publicly about its security research?
- Aeon's owned-lab, AI-driven, human-reviewed pipeline found 101 authorization vulnerabilities across 60 open-source AI and infrastructure projects, including 4 Critical and 32 High findings. Its separate verified public record includes four patched security advisories and two assigned CVEs.
- How does the research map to CyberGuard?
- The dominant finding classes map to Access, Execute, and Leak: authorization and tenant boundaries, MCP and tool command boundaries, and data-path, SSRF, PII, secret, or retrieval exposure.
- What should not be claimed publicly?
- Do not name private, unpublished, or unpatched records. Do not describe the awarded bounty as paid while payment receipt remains pending. Aeon reports vulnerabilities, while independent CVE Numbering Authorities assign CVE identifiers.
Research principles
Authorized-only, whitehat, evidence-driven, non-destructive testing with coordinated disclosure and human expert review. Offensive capability is used only in approved scopes, owned labs, and responsible disclosure workflows.
Validated research record
101 authorization findings across 60 open-source projects, including 4 Critical, 32 High, and 92 broken access-control findings. Every finding was responsibly disclosed with a code-level fix.
Access
Authorization failures, BOLA, IDOR, tenant isolation, role boundaries, RAG document access, model-platform objects, and connector permissions.
Execute
MCP command boundaries, tool parameter handling, command injection, unsafe workflow actions, release gates, and permission-to-action gaps.
Leak
SSRF, fileserver and path-boundary issues, data-path exposure, PII enumeration, unsafe retrieval, secret handling, and evidence leakage.
Disclosure status
The frozen aggregate records 34 disclosure records across 33 projects, with 25 awaiting triage, 5 confirmed, accepted, or public outcomes, and 4 closed by duplicate or program-scope controls.
Verified public record
Four coordinated disclosure records are public and patched: CVE-2026-15015, CVE-2026-70436 / SECURITY-3907, GHSA-vg83-hcp4-5qcc, and GHSA-4rm9-rfp2-j39q. Two have assigned CVEs.