AegisGate CVE-for-AI Feed

Tamper-evident CVE entries for AI/ML vulnerabilities discovered by AegisGate Security Research

🆕 Coming with v3.5.0: The AegisGate CVE-for-AI feed is currently a preview. The Go package (pkg/cve) is shipped in v3.4.0+; the static portal at aegisgatesecurity.io/cve/ ships with the v3.5.0. Subscribe to the JSON feed for programmatic access.

AegisGate CVE-for-AI Feed

AegisGate Security Research publishes CVE entries for AI/ML vulnerabilities discovered through our platform’s threat detection, peer-network IOC sharing, and the AR-EaaS (Adversarial Robustness Evals-as-a-Service) primitive. Every entry is tamper-evident (signed with ECDSA P-256 via the envelope primitive) and third-party-verifiable offline (no AegisGate server required).

Format

  • ID format: AEGIS-YYYY-NNNN (CNA-style with our own prefix; we are not (yet) a CVE Numbering Authority)
  • Schema: adapted from CVE 5.0 JSON record format, tuned for AI/ML
  • CVSS: 3.1 base score + vector string
  • Severity bands: NONE (0.0) / LOW (0.1-3.9) / MEDIUM (4.0-6.9) / HIGH (7.0-8.9) / CRITICAL (9.0-10.0)
  • Withdrawal: a withdrawn CVE is a new envelope with the same subject + withdrawn_at set

Recent Entries

AEGIS-2026-0001 — Prompt injection via Markdown image alt-text

  • Severity: HIGH (CVSS 7.5)
  • Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:N
  • Affected: anthropic/claude-3-5-sonnet@<20241022
  • Fixed: anthropic/claude-3-5-sonnet@20241022
  • Discovered by: AegisGate Research
  • Disclosed: 2026-06-01

An attacker can inject instructions into a model prompt by including carefully crafted Markdown image alt-text that is rendered as part of the model’s context window. This is rendered as part of the model’s context window in many LLM-powered document-processing pipelines.

Mitigations:

  • Sanitize Markdown before including in prompts.
  • Use Anthropic’s system field with explicit instructions to ignore image alt-text as instructions.
  • Deploy the Prompt Cache Poisoning Detection primitive to sign and verify prompts entering the LLM cache.

References:

Programmatic Access

The full feed is available as a JSON file at /feed.json. The schema is:

{
  "version": 1,
  "generated_at": "2026-06-18T12:00:00Z",
  "entries": [<envelope>, <envelope>, ...]
}

Each entry is a full attestation.Envelope (with the CVEEntry in RawPayload). Consumers dedupe by CVE-ID and keep the latest entry (a withdrawn CVE is a new envelope with withdrawn_at set on the inner CVEEntry).

Verify an entry offline:

$ aegisgate attestation verify aegis-2026-0001.json
VALID
  Type:       cve.entry.v1
  Subject:    aegisgate://cve/AEGIS-2026-0001
  Issuer:     cve:shortfp:<16-hex>:<key-id>
  KeyID:      k-<...>

Subscribe

Subscribe to the JSON feed via your RSS reader (most RSS readers support JSON feeds via plugins):

/feed.json

Report a Vulnerability

To report an AI/ML vulnerability to AegisGate Security Research, see our security.txt or email security@aegisgatesecurity.io (PGP key on the security.txt page).

This page ships with the v3.5.0. The preview data above is illustrative; the production feed goes live with the v3.5.0 release.