AI Agent Skill Supply Chain Poisoning & Integrity Simulator

Verified Tool Registry / Pinned Node

Active AI Agent Cluster

Poisoned Endpoint / Hijacked Node

Live Agent Skill Call Log 2,450 calls/sec

[00:00:01] AGENT-FLEET-A fetched skill web_search (v1.0.2) from registry.io

⚡ Attack Vectors

🔒 Integrity Verification Protocol

Cryptographic Skill Pinning

Lock target endpoints to strict SHA-256 content hashes

Hardware Enclave Signature

Verify Ledger enclave attestation before execution

📊 Real-Time Security Telemetry

Simulated Agent Fleet

26,000

Compromised Agents

Blocked Attacks

26,000

Integrity Score

100%

Agent Fleet Scaling 26,000 Agents

🔍 Security Audit Proof

Total Agents: 26000

Compromised Pinned: 0

Compromised Unpinned: 26000

Blocked Attacks: 26000

Integrity Status: SECURED

📄 Skill Lockfile & Policy Export

{ "version": "1.0.0", "pinned_skills": { "web_search": { "url": "https://skills.registry.io/v1/search", "hash": "a8f3b2c1d4e5f607182930415263748596a7b8c9d0e1f2a3b4c5d6e7f8a9b0c1", "enclave_signed": true } } }

Skill supply-chain simulation: pinning labels versus verification

Your browser does not support video.

Read the explanation

The saved skill-supply-chain simulator applies two state switches: whether the upstream URL is hijacked and whether protection mode is cryptographic pinning. Under hijack with unpinned mode it sets compromised count to the entire configured fleet, twenty-six thousand by default. With pinning it assigns zero compromised and all fleet members blocked. The bars use one pixel per fifty configured agents. This is a binary simulation rule, not an observed compromise rate or real fleet audit. It does not fetch code, compute a digest, verify a signature or operate a hardware enclave. The lockfile view marks verified from the protection-mode flag; that Boolean label cannot establish actual cryptographic verification. When there is no hijack, the simulator still sets blocked attacks equal to fleet size. A positive blocked count therefore does not prove any attack occurred. Changing fleet size rescales assigned counts. At the default twenty-six thousand agents, the bar is one hundred thirty pixels; at a hundred thousand, it is five hundred pixels. Both use one pixel per two hundred configured agents, so the geometry faithfully compares counts. The graph always contains twelve representative cluster nodes, not twenty-six thousand individual workers. All twelve clusters reference the search registry and even-numbered clusters also reference the transaction registry. These are diagram elements and assigned state values, not active processes, observed attacks or backend calls. The hidden semantic proof fields always assign zero to compromised pinned and full fleet size to compromised unpinned, independent of the current attack. They represent fixed hypothetical alternatives rather than two measured cohorts. The visible integrity score is also binary: one hundred percent when compromised count is zero and zero percent otherwise. The exported policy and lockfile serialize those states and fixed skill hashes; exporting does not enforce a trust policy or validate executable artifacts. In a real implementation, trusted digest provenance and verification before loading would need independent evidence. This explanation describes saved source and native controls without contacting registries, loading skills, using credentials or publishing changes.