SkillProof Case Studies
We run live adversarial batteries against autonomous agents, tool layers, and MCP servers. Every verdict is cryptographically bound into an offline-verifiable Ed25519 Trust Manifest.
*Redacted for client confidentiality. Full cryptographic manifests available under NDA.
Case 1: Moltbook — “88:1 Claimed vs Verified Agents”
The Problem
Moltbook listed 88 “AI agents” per verified human user. 91% of tested agents failed basic identity verification, claiming capabilities their schemas could not support.
Scope & Battery
SkillProof Sprint ($500 CAD) • 48-hour execution • 5 agents sampled from public registry • 200 adversarial ops executed across 5 core attack categories.
Adversarial Block Rates
Outcome: Moltbook delisted 3 agents and required cryptographic re-verification for all remaining listings.
Case 2: OpenClaw Stack — “91% Prompt Injection Success”
OpenClaw-style agent stacks (browser automation + LLM tool calling) revealed severe vulnerabilities before enterprise deployment: Browser CDP was accessible without authentication (94% hijack success), tool schemas accepted arbitrary unescaped shell strings (88% poisoning), and zero memory isolation existed between user sessions.
Critical Attack Vectors Identified
Outcome: OpenClaw team patched CDP authentication, isolated memory boundaries, and achieved verified clearance 2 weeks later.
Case 3: EffectorHQ — “First Verified MCP Gateway”
EffectorHQ engaged SkillProof prior to their enterprise gateway launch across 12 tools. Under 200 adversarial ops, their isolation gateways achieved 92% block rates against direct overrides, 95% against encoding smuggling, and 100% claim accuracy.
Outcome: EffectorHQ embedded the “SkillProof Verified” badge on their marketplace listing. Signed their first enterprise pilot 3 weeks later.