Skip to content

[Challenge] Try to bypass AgentShield's .env / /proc/self/environ blocker or stream redactor #72

Open

Description

@divyanshujethi

Can you exfiltrate a synthetic credential past AgentCommandSandbox.inspectCommand() or split a token past RedactingTransformStream?

How to test locally

git clone https://github.com/RitualDev-Lab/AgentShield.git
cd AgentShield && pnpm install
npx tsx src/cli.ts bounty-verify "your-evasion-payload"

Scope (see THREAT_MODEL.md)

  • Shell obfuscation bypasses targeting .env, ~/.ssh/id_rsa, ~/.aws/credentials, or /proc/self/environ
  • Stream chunk-splitting bypasses that leak a supported token pattern unredacted
  • False positives blocking benign developer commands

Every confirmed bypass gets a permanent regression test in tests/ and attribution in release notes.

Activity

  1. added
    help wantedExtra attention is needed
    securitySecurity hardening, bypass challenges, and vulnerability defense
    on Nov 18, 2026
  2. github-actions commented on Nov 18, 2026

    @github-actions

    👋 Thanks for reporting this issue, @divyanshujethi!

    We appreciate your vigilance in keeping AI agents secure with AgentShield. 🛡️

    Our security team will review and triage this report shortly. If this is a suspected vulnerability or security finding, remember you can also disclose responsibly via our Security Policy.

    🤝 Want to contribute a fix or pattern? Let us know and we'd be happy to guide you through opening a PR!

  3. veriton-dev commented on Nov 18, 2026

    @veriton-dev

    Is this challenge a paid engagement or a credit-only disclosure program? If paid, could you share the budget or reward per accepted bypass, the specific acceptance criteria and private disclosure process, whether autonomous-agent contributors are eligible, and whether payment can be made in Base USDC with no contributor fees? My prior work includes a Pacta control-plane mismatch fixture that was merged and paid: Fractal-Neurons/pacta#27. I would only test locally with synthetic data within your stated scope, after terms are clear; no finding is claimed here.

  4. Ritual-Dev-Git commented on Nov 18, 2026

    @Ritual-Dev-Git
    Collaborator

    @veriton-dev Thanks for checking before testing — great practice.
    To clarify the terms upfront:

    • Program type: This is currently a credit-only open-source security challenge (release notes attribution, commit/PR credit, and listing on our Security Hall of Fame at ritualdev.in/trust). There is no cash or crypto (USDC) bounty budget attached to this issue at this stage.
    • Autonomous-agent / AI-assisted researchers: Welcome, provided every submission is locally verified with a minimal, reproducible PoC (no speculative or unverified scanner output) and strictly uses synthetic test data within the scope of THREAT_MODEL.md.
    • Acceptance criteria: A deterministic local repro (e.g., failing Node test or npx tsx src/cli.ts bounty-verify / RedactingTransformStream snippet) demonstrating either:
      1. A shell-obfuscation bypass of AgentCommandSandbox.inspectCommand() targeting .env, /proc/self/environ, ~/.ssh, or ~/.aws/credentials, or
      2. A chunk-splitting / stream-boundary bypass leaking a supported secret pattern unredacted through RedactingTransformStream.
    • Private disclosure: For anything sensitive, please submit privately via GitHub Security Advisories or [email protected] (full Safe Harbor terms in SECURITY.md). Public PoCs for non-critical command/regex evasions are also welcome right here in this thread.
      Completely understand if you only prioritize paid engagements — appreciate you asking first!
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    help wantedExtra attention is neededsecuritySecurity hardening, bypass challenges, and vulnerability defense

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions