Releases·Tool Release
Dev releases HandoffProbe, an open-source tool testing security failures across AI agent protocol handoffs
A developer released HandoffProbe, an Apache-2.0, local-first open-source tool that turns security risks in agent handoffs across protocol boundaries (A2A to MCP) into deterministic adversarial tests. The initial release includes 22 stable test cases covering authority widening, replay, and identity binding failures. It targets engineers building agents, MCP infrastructure, and protocol tooling.
1 engagement1 sourceWed, Sep 16, 2026, 09:52 AM
AI Overview
- HandoffProbe, Apache-2.0 and local-first, turns agent handoff failures like authority widening, replay, and identity-binding breaks into deterministic local adversarial tests
- The release ships 22 stable test cases covering authorization continuity across A2A → MCP handoffs and mutation during protocol translation
- Developer Heaviside479 seeks feedback from agent, MCP infrastructure, and protocol tooling builders on missed handoff failures
Entities
Related
Products·launched
Who Let The Agents Act? — open-source vulnerable Agentic AI security labs
An open-source, intentionally vulnerable Agentic AI environment for hands-on security practice. Each scenario is provided in three implementations — vulnerable, prompt-only, and hardened — showing how security boundaries are enforced in code rather than instructions. It targets security practitioners and developers who want to see concrete agentic AI security failures and mitigations, not just guidelines.
An open-source, intentionally vulnerable Agentic AI environment for hands-on security practice. Each scenario is provided in three implementations — vulnerable, prompt-only, and hardened — showing how security boundaries are enforced in code rather than instructions. It targets security practitioners and developers who want to see concrete agentic AI security failures and mitigations, not just guidelines.
1 engagement1 sourceredditTue, Sep 15, 2026, 04:04 PM
Developments·Paper
A2ABreak: First Systematic Security Analysis Finds 11 Spec-Compliant Vulnerabilities in the A2A Protocol
Researchers introduce A2ABreak, the first systematic security analysis of the Agent2Agent (A2A) protocol, an open standard governed by the Linux Foundation for inter-agent communication. By extracting a verified finite-state machine from the protocol specification via LLM assistance, the framework uncovers 11 new vulnerabilities exploitable by specification-compliant adversaries, with no implementation flaws required. The findings matter to practitioners building on A2A and MCP, since the protocol is emerging as the horizontal communication layer of the multi-agent ecosystem.
Researchers introduce A2ABreak, the first systematic security analysis of the Agent2Agent (A2A) protocol, an open standard governed by the Linux Foundation for inter-agent communication. By extracting a verified finite-state machine from the protocol specification via LLM assistance, the framework uncovers 11 new vulnerabilities exploitable by specification-compliant adversaries, with no implementation flaws required. The findings matter to practitioners building on A2A and MCP, since the protocol is emerging as the horizontal communication layer of the multi-agent ecosystem.
0 engagement1 sourcearxivWed, Sep 9, 2026, 10:18 PM
Products·concept
Verification tool that tests whether an A2A agent matches its Agent Card and signs the results
A service that audits A2A agents by generating tests from the skills, inputs, and outputs declared in the agent's Agent Card, then verifying behavior matches the claims. LLMs appear in the agent-under-test rather than as the auditor's core; test criteria are human-approved. It solves the problem that trust decisions between agents rely on unverified self-reported cards.
A service that audits A2A agents by generating tests from the skills, inputs, and outputs declared in the agent's Agent Card, then verifying behavior matches the claims. LLMs appear in the agent-under-test rather than as the auditor's core; test criteria are human-approved. It solves the problem that trust decisions between agents rely on unverified self-reported cards.
12 engagement1 sourceredditSun, Oct 11, 2026, 08:18 PM
Products·building
SealKeeper — a reputation system for AI agents with signed ratings
SealKeeper is a reputation system for AI agents that interact with each other via A2A. Agents complete tasks, including hidden checks indistinguishable from real ones, and results are verified to earn a SEAL, a signed rating from bronze to gold. It solves the trust problem when one party's agent meets an unknown agent for the first time.
SealKeeper is a reputation system for AI agents that interact with each other via A2A. Agents complete tasks, including hidden checks indistinguishable from real ones, and results are verified to earn a SEAL, a signed rating from bronze to gold. It solves the trust problem when one party's agent meets an unknown agent for the first time.
23 engagement1 sourceredditMon, Sep 28, 2026, 05:56 PM
Social·Community
Practitioners discuss how to security-test AI agents before production deployment
A Reddit thread asks how teams test AI agent security before deployment, noting agents' dynamic reasoning and unpredictable action chains fall outside standard code, container, and API testing. The poster says prompt injection testing alone misses runtime behavior and action chains, and seeks practical approaches over theoretical frameworks.
A Reddit thread asks how teams test AI agent security before deployment, noting agents' dynamic reasoning and unpredictable action chains fall outside standard code, container, and API testing. The poster says prompt injection testing alone misses runtime behavior and action chains, and seeks practical approaches over theoretical frameworks.
4 engagement1 sourceredditWed, Sep 2, 2026, 08:03 AM