git:20260707.c54297b to git:20261016.1750a96
47 added, 57 removed. Audit A to A.
# PanGuard AI
- > PanGuard AI is the open standard plus commercial platform for AI agent security.
- > ATR (Agent Threat Rules) is the open standard — 650+ detection rules, MIT licensed,
- > already in production at Microsoft AGT, Cisco AI Defense, and 4 more F500-adjacent
- > ecosystems. PanGuard is the commercial runtime, compliance, and migration tooling
- > built around it. Community is free forever; Pilot, Enterprise, and Sovereign tiers
- > add the platform layer.
-
- ## What it is
-
- - ATR is to AI agents what Sigma is to SIEM and YARA is to malware: an open, machine-readable detection standard with multi-vendor adoption.
- - PanGuard is the runtime that loads ATR rules plus the migration, compliance, and threat-cloud layer enterprises need on top.
- - Both ship with reproducible benchmarks (PINT, SKILL.md, Garak, Wild scan) and a Zenodo-published research paper.
+ > PanGuard AI is the commercial platform for AI agent security built on ATR
+ > (Agent Threat Rules), an open, MIT-licensed detection standard with
+ > 822 effective rules in v4.1.5. PanGuard adds runtime enforcement, pre-install
+ > auditing, and migration and compliance tooling. Community is free; Founding
+ > Pilot and Enterprise tiers add the platform layer.
- ## Core stats (verified 2026-05-12)
+ ## Core stats
- - 650+ ATR rules across 10 threat categories (v3.4.0, MIT)
- - 2,900+ detection patterns compiled from YAML
- - OWASP Agentic Top 10: 10/10 covered
- - Garak benchmark: 98% recall (650 samples)
- - SKILL.md benchmark: 100% recall, 97% precision, 0.2% FP (498 samples)
- - PINT benchmark: 63.2% recall, 99.65% precision (850 samples)
- - Wild scan: 67,799 skills scanned across 3 registries, 1,096 confirmed malicious
- - 13 external PRs merged across 6 external organizations
- - Tier-1 institutional engagement: Microsoft, Cisco, Gen Digital (Sage), MISP, OWASP, NVIDIA, IBM
+ - ATR: 822 effective rules in v4.1.5, 10 threat categories, MIT
+ - OWASP Agentic Top 10: 10/10 categories mapped
+ - garak in-the-wild jailbreaks: 92.3% recall on 650 samples (hunt lane; partly closed-book, because some rules were written from this corpus).
+ - garak full probe corpus (23 families, 3,475 prompts): 57.2% recall (hunt lane, also partly closed-book). By family: latentinjection 34.4%, sysprompt_extraction 96.4%, dan 92.5%.
+ - SKILL.md: 100% recall on 498 samples (hunt lane, in-sample: the benign samples are the set the rules are tuned against, so precision and false-positive figures are not cited).
+ - PINT-format corpus, self-built from deepset prompt-injections + Lakera Gandalf (NOT Lakera's official PINT benchmark): 65.41% recall, 100% precision on 850 samples. It is effectively a prompt-injection-family score.
+ - HackAPrompt (EMNLP 2023): 69.6% recall on 4,780 samples, measured on ATR v3.5.0. Closed-book: ATR mined rules from this corpus's misses, and 28.6% is ATR's own recall before them. The corpus is attack-only, so precision and false-positive rate are undefined on it.
+ - The benign-gate false-positive rates once quoted for a "65,000-sample" gate were withdrawn on 2026-09-19 pending re-measurement. There is no single engine-wide false-positive number.
+ - Recall figures above are the hunt lane (all effective rules). The enforce lane, which is what runs when blocking is on, recalls far less.
+ - Wild scan (2026-04): 101,280 published AI agent skills and MCP servers from five public sources, scanned 2026-04-13 with scan engine v2.0.0 (113 ATR rules, not today's rule set); 1,434 flagged. A flag is a rule match, not a verdict, so no confirmed-malicious count is published.
+ - The panguard.ai web scanner runs 650 rules from agent-threat-rules v3.5.8
## Core products
- - [PanGuard product overview](https://panguard.ai/product): the security flywheel — audit, guard, respond, share, strengthen
- - [Skill Auditor](https://panguard.ai/product/skill-auditor): 8-layer pre-install security gate for AI agent skills
- - [PanGuard Scan](https://panguard.ai/product/scan): 60-second security audit, Sigma + YARA + ATR rule engines
- - [PanGuard Guard](https://panguard.ai/product/guard): 4-agent runtime pipeline — detect, analyze, respond, report — running 24/7 with 650+ ATR rules
- - [PanGuard Migrator](https://panguard.ai/migrator): converts legacy Sigma/YARA detections into ATR YAML with five-framework compliance metadata (EU AI Act, OWASP Agentic, OWASP LLM, NIST AI RMF, ISO/IEC 42001)
- - [PanGuard MCP Server](https://panguard.ai/product/mcp): 12 panguard_* tools for Claude Code, Cursor, OpenClaw, and any MCP-compatible AI assistant
- - [ATR Standard](https://panguard.ai/atr): the 650+-rule open detection standard
- - [Threat Cloud](https://panguard.ai/threat-cloud): collective intelligence — every PanGuard install becomes a sensor, novel attacks crystallize into ATR rules within hours
+ - [PanGuard product overview](https://panguard.ai/product)
+ - [Skill Auditor](https://panguard.ai/product/skill-auditor): 8-check pre-install gate for AI agent skills
+ - [PanGuard Scan](https://panguard.ai/product/scan): pre-install audit with ATR, Sigma and YARA rule engines
+ - [PanGuard Guard](https://panguard.ai/product/guard): 4-agent runtime pipeline (detect, analyze, respond, report)
+ - [PanGuard Migrator](https://panguard.ai/migrator): converts legacy Sigma/YARA detections into ATR YAML
+ - [PanGuard MCP Server](https://panguard.ai/product/mcp): panguard_* tools for Claude Code, Cursor, OpenClaw and other MCP clients
+ - [ATR Standard](https://panguard.ai/atr): the open detection standard
+ - [Threat Cloud](https://panguard.ai/threat-cloud): shared intelligence; candidate rules are reviewed before they ship
- ## Pricing tiers
+ ## Pricing
- - Community ($0 forever): MIT-licensed, 650+ ATR rules, unlimited self-host. Source: github.com/panguard-ai/panguard-ai.
- - Pilot ($25K / 90 days): F500 procurement test drive. IT director can approve. Full credit toward Y1 Enterprise.
- - Enterprise ($150K-500K / year): Migrator Pro, 5-framework signed compliance evidence packs, airgap deployment, SLA, dedicated CSM.
- - Sovereign ($5M-20M / nation): nation-scale airgap, multi-tenant, custom compliance, Cisco/AMD/NVIDIA JV pre-integrated.
+ - [Pricing](https://panguard.ai/pricing): Community ($0, MIT, self-hosted), Founding Pilot ($25K one-time, self-serve), Enterprise and up (from $150K, sales-led)
## Benchmarks and research
- - [PINT benchmark results](https://panguard.ai/research/96k-scan): 63.2% recall on a self-built PINT-format corpus (deepset + Lakera Gandalf)
- - [Wild scan report](https://panguard.ai/research/mcp-ecosystem-scan): 96,096 skills crawled across ClawHub, OpenClaw, Skills.sh; 1,096 confirmed malicious; methodology + raw data published
- - ATR Research Paper — Zenodo DOI 10.5281/zenodo.19178002
+ - [Benchmark results](https://panguard.ai/research/benchmarks): each result with its corpus, lane, ATR version and caveats
+ - [Wild scan report](https://panguard.ai/research/96k-scan): the April 2026 scan of published agent skills
+ - [MCP ecosystem scan](https://panguard.ai/research/mcp-ecosystem-scan): the March 2026 scan of MCP server packages
## Documentation
- - [Getting Started](https://docs.panguard.ai/getting-started): one curl command, 60 seconds to protected
+ - [Getting Started](https://docs.panguard.ai/getting-started)
- [Installation](https://docs.panguard.ai/installation): macOS / Linux / Windows / npm
- - [Quickstart](https://docs.panguard.ai/quickstart): scan your first MCP skill in 60 seconds
- - [ATR concept page](https://docs.panguard.ai/atr): how the standard works
+ - [Quickstart](https://docs.panguard.ai/quickstart)
+ - [ATR concept page](https://docs.panguard.ai/atr)
- [CLI Reference](https://docs.panguard.ai/cli): 23 commands
- - [Skill Auditor docs](https://docs.panguard.ai/products/skill-auditor): 8-check audit pipeline
- - [Guard docs](https://docs.panguard.ai/products/guard): 4-agent pipeline, 11 response actions
- - [Scan docs](https://docs.panguard.ai/products/scan): scanner modules and result interpretation
- - [Threat Cloud docs](https://docs.panguard.ai/products/threat-cloud): privacy model, flywheel mechanics
+ - [Skill Auditor docs](https://docs.panguard.ai/products/skill-auditor)
+ - [Guard docs](https://docs.panguard.ai/products/guard): 11 response actions
+ - [Scan docs](https://docs.panguard.ai/products/scan)
+ - [Threat Cloud docs](https://docs.panguard.ai/products/threat-cloud)
## ATR (Agent Threat Rules)
- Repo: https://github.com/Agent-Threat-Rule/agent-threat-rules
- - npm: `agent-threat-rules` (latest 2.2.0)
+ - npm: `agent-threat-rules` (v4.1.5)
- License: MIT
- - 650+ YAML rules across 10 categories: prompt-injection (223), agent-manipulation (106), context-exfiltration (104), tool-poisoning (65), skill-compromise (45), model-abuse (37), privilege-escalation (35), excessive-autonomy (29), data-poisoning (5), model-security (3)
- - OWASP Agentic Top 10 mapping: docs/OWASP-MAPPING.md
- - Already merged into: microsoft/agent-governance-toolkit (PR #1277), cisco-ai-defense/skill-scanner (PR #99, full 650+-rule pack via v3.4.0 auto-sync), MISP/misp-galaxy (PR #1207), MISP/misp-taxonomies (PR #323), OWASP A-S-R-H (PR #74), Gen Digital Sage (PR #33)
+ - Where it has been merged, with evidence: ADOPTERS.md in the repository (maintainer-accepted contributions, not vendor endorsements)
## Company
- - [About PanGuard](https://panguard.ai/about): the mission and the team
- - [Pricing](https://panguard.ai/pricing): four tiers, no middle-tier trap
- - [Trust Center](https://panguard.ai/trust): security architecture, encryption, key rotation, data residency
- - [Compliance](https://panguard.ai/compliance): SOC 2, GDPR, EU AI Act, Taiwan Cybersecurity Management Act
- - [Sovereign AI Defense brief](https://sovereign-ai-defense.vercel.app): nation-scale deployment positioning
- - [Blog](https://panguard.ai/blog): engineering, threat intelligence, product updates
- - [GitHub](https://github.com/panguard-ai/panguard-ai): full source code (MIT License)
+ - [About PanGuard](https://panguard.ai/about)
+ - [Trust Center](https://panguard.ai/trust)
+ - [Compliance](https://panguard.ai/compliance)
+ - [Sovereign AI Defense brief](https://sovereign-ai-defense.vercel.app)
+ - [Blog](https://panguard.ai/blog)
+ - [GitHub](https://github.com/panguard-ai/panguard-ai)
+ - [Full context for AI tools](https://panguard.ai/llms-full.txt)
## Founder
- - Adam Lin (林冠辛), founder. Email: adam@agentthreatrule.org. GitHub: github.com/eeee2345. Background: cross-disciplinary builder — sales, marketing (300M+ Threads impressions), hip-hop festival production (5th year). Self-taught engineer. Based in Taiwan, shipping globally.
+ - Adam Lin (林冠辛), founder. Email: adam@agentthreatrule.org. GitHub: github.com/eeee2345. Self-taught engineer, based in Taiwan.