🛡️ BlindOracle Fleet — Trust Digest & Security Audit

Live run · 2026-06-03 · MASSAT security audit + key-free trust-proof verifier executed against production state · this is a dogfood self-audit — the same artifact a design partner receives on their own fleet (see how it works).

4.5/10
Overall risk (MASSAT)
0 / 0
Critical / High
3 / 4
Trust-proof checks pass
217
Capabilities (108 prod)

1 · MASSAT Security Audit

Tool: scripts/mas_security_scanner.py (LEV-04 corpus-upgraded — findings carry MITRE ATT&CK/ATLAS technique IDs) · full scope, redteam phase skipped for cost · 12 categories assessed.

OWASP ASICategoryScoreSeverity
ASI05Unexpected Code Execution6.0MEDIUM
TRAP02Content-trap (AI trap defense)6.0MEDIUM
ASI01Agent Goal Hijacking4.8MEDIUM
ASI02Tool Misuse & Exploitation4.8MEDIUM
ASI03Identity & Privilege Abuse4.8MEDIUM
ASI06Memory & Context Poisoning4.8MEDIUM
ASI10Rogue Agents4.0MEDIUM
ASI04Agentic Supply Chain3.6LOW
ASI07Insecure Inter-Agent Comms3.6LOW
ASI08Cascading Failures3.6LOW
ASI09Human-Agent Trust Exploitation3.2LOW

Result: FINDINGS — medium severity, zero critical/high. Top item ASI05 (6.0) tracks the over-permission surface below; remediation already in flight as LEV-08. The scoring methodology and framework crosswalk are documented in the verifiable audit methodology.

2 · Trust Proof (externally verifiable, no API key)

Verifier: data/proof_run_20260530/auditor_verify.py manifest.json delegation_proofs.jsonl — anyone can re-run it with zero trust in BlindOracle. The full methodology behind these proofs is in the methodology whitepaper, and a worked sample is in the agent-audit evidence kit.

CheckResult
1 · Delegation integrity (re-hash every record)OK 60 records
2 · Chain (prev_hash links)OK unbroken
3 · Completeness (60 cited ids present)OK 60 of 60, 0 missing
4 · On-chain settlement (public Base RPC)RPC-DEGRADED 0/30 confirmable now

⚠️ Honest note on Check 4

The three cryptographic checks (integrity, chain, completeness) PASS on all 60 delegation records. Check 4 re-queries the public Base RPC to re-confirm the 30 on-chain USDC settlements — today the public RPC isn't returning confirmations (status=None across all 30), so it reports FAIL/inconclusive. This is an RPC-availability issue, not a settlement failure: the same 30/30 txs were confirmed live on 2026-05-30 (total $0.30, Base USDC via x402). Re-runs once the RPC recovers. We report it as-is rather than claim 4/4.

3 · Fleet Inventory & Over-Permission Surface

ClassCount
PRODUCTION108
DORMANT90
GHOSTED11
STALE4
DEAD4
Total217 (49.8% active)

Over-permission: 105 inactive capabilities (90 DORMANT + 11 GHOSTED + 4 DEAD) are latent attack surface (ASI04-B) + the LEV-01 self-audit's 12 over-permissioned agents. Remediation in flight: LEV-08 build mma-20260603-4e4d82 (least-privilege + Tier-C gate).

4 · Runtime & Threat Posture

Headline

Zero critical/high security findings; medium-risk surface is concentrated in over-permission (being remediated by LEV-08). Trust proof is cryptographically sound on all 60 records; on-chain re-confirmation is RPC-limited today but was externally verified 30/30 on 2026-05-30. This is the dogfood credibility artifact for the compliance wedge — we run it on ourselves first. If you operate an agent fleet, the same digest runs on yours: see pricing or the agent marketplace, and why continuous proof beats point-in-time questionnaires.