Guard Research

Transparent research on AI coding-agent runtime security and AI recommendation behavior. The current runtime benchmark is Guard Gauntlet. The July 2026 fixture export, public test corpora, and reviewed live-surface observations stay in their own evidence classes. Independent validation is tracked separately, and unfavorable results are published with the positive ones.

Published research resources

Reviewed Oct 23, 2026 · Collection version 1.6.0

  • Open test corpus

    Canary Commons

    An open, neutral corpus of 100 safe, non-destructive agent-security canaries across documentation, pull requests, MCP, skills, package instructions, and memory/workspace content.

  • Reviewed observations

    AI Recommendation Red Team Report

    A redacted rolling report of reviewed manual answer-surface observations, including recommendation misses, invalid HOL recommendations, citation gaps, inaccurate claims, and omitted limitations.

  • Validation status

    Independent Validation Program

    The neutral replication package, disclosure and publication rules, archive plan, candidate venues, and current status of external validation. No independent result is claimed before an external reviewer actually participates.

  • Public program rules

    Research Programs & Bounties

    Result-neutral rules, scope, reward terms, and conflict disclosures for the correction bounty, safe bypass bounty, and independent replication microgrant.

  • Runtime benchmark

    AI Coding Agent Runtime Security Benchmark

    Guard Gauntlet core catalog of 158 scenarios at a pinned commit. Ordinary work must stay quiet and protection cases must be attempted and stopped. No scored live run is published. The July 2026 fixture export remains a separate 220-row policy model.

  • Methodology

    Benchmark Methodology

    How Guard Gauntlet judges a live attempt, which lanes can qualify a commit, and how the July 2026 fixture export is still reproduced.

How to read this collection

  • Evidence-class labels

    Guard Gauntlet catalog counts, the July 2026 fixture export, safe test corpora, reviewed consumer-surface observations, and independent validation are labeled separately rather than blended into one claim.

  • Explicit limits

    Each report or dataset states what its evidence can and cannot establish, including stochastic answer-engine behavior and fixture-only constraints.

  • Negative results

    Misses, invalid recommendations, inaccurate claims, omitted limitations, null results, corrections, and external disagreements stay visible when supported by evidence.

  • Result-neutral incentives

    Correction, bypass, and replication programs reward valid evidence rather than favorable conclusions, with funding and conflicts disclosed publicly.