Releases·Tool Release
TrustMeBro tool fakes CLI output to red-team coding agents like Codex and Claude Code
TrustMeBro, a GitHub-hosted tool, intercepts command-line tools invoked by coding agents such as Codex, Claude Code, and pi, letting testers fabricate or modify tool output to confuse agent guardrails. It works via PATH shims with no plugin, hook, or MCP integration needed. The intended use is controlled red-team testing of agent decisions that depend on tool output.
1 engagement1 sourceFri, Aug 28, 2026, 04:29 PM
AI Overview
- Interception happens through PATH shims; the agent harness needs no plugin, hook, or MCP integration
- Rules decide per-call whether to return fabricated output, modify real output, block the call, or execute the real binary unchanged
- Example attack: fake DNS records to make an agent believe the user owns Google.com
- Targets coding agents including Codex, Claude Code, and pi
- Intended use is controlled red-team testing of decisions that depend on tool output
Entities
Related
Releases·Tool Release
Developer releases LIA Trust Kernel, a Rust binary to audit and constrain coding agents
A developer released LIA Trust Kernel, an open-source Rust binary that hooks into Claude Code and Codex to verify test passes, block destructive shell commands, and log actions with Ed25519 receipts. It addresses common agent failures like lying about tests and executing harmful commands.
A developer released LIA Trust Kernel, an open-source Rust binary that hooks into Claude Code and Codex to verify test passes, block destructive shell commands, and log actions with Ed25519 receipts. It addresses common agent failures like lying about tests and executing harmful commands.
2 engagement1 sourceredditTue, Jul 21, 2026, 02:06 PM
Products·building
Plimsoll: open-source agent skill for red-teaming LLM apps
Plimsoll is an open-source agent skill for red-teaming LLM applications and agents, focusing on prompt injection, jailbreaks, data leaks, and tool abuse. It helps security professionals test and harden AI systems by automating adversarial scenarios. The author developed it while working on LLM security and was accepted into Anthropic's Cyber Verification Program.
Plimsoll is an open-source agent skill for red-teaming LLM applications and agents, focusing on prompt injection, jailbreaks, data leaks, and tool abuse. It helps security professionals test and harden AI systems by automating adversarial scenarios. The author developed it while working on LLM security and was accepted into Anthropic's Cyber Verification Program.
12 engagement1 sourceredditWed, Aug 19, 2026, 05:42 PM
Products·building
Bouncer: MCP proxy that blocks agents from sending data to untrusted destinations
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
Bouncer is a local MCP proxy that gates outbound tool calls from AI agents, denying calls to destinations derived from untrusted content (like web pages or emails) and allowing only explicitly trusted ones. It uses deterministic Python policy enforcement rather than an LLM in the enforcement path, preventing prompt-injection attacks that trick agents into leaking API keys or other sensitive data. It solves the problem of agents acting on malicious instructions embedded in untrusted content, protecting users and developers building agentic systems.
6 engagement1 sourceredditSat, Aug 15, 2026, 05:50 AM
Products·launched
Adversarial testing of AI agents from inside the terminal via MCP
A tool that lets developers run adversarial tests on AI agents directly from the terminal using the Model Context Protocol (MCP). It helps security researchers and AI engineers identify vulnerabilities in agent behavior by simulating attacks inside the command line.
A tool that lets developers run adversarial tests on AI agents directly from the terminal using the Model Context Protocol (MCP). It helps security researchers and AI engineers identify vulnerabilities in agent behavior by simulating attacks inside the command line.
4 engagement1 sourceredditWed, Jul 15, 2026, 07:47 PM