We break things.
Reproducible proofs of concept for prompt-injection chains, tool and function-call abuse, confused-deputy attacks and data exfiltration through agents. If it runs autonomously and touches a tool, we treat it as in-scope.
// offensive security for AI agents
Breaking autonomous agentsso you can harden them.
secagentlabs is an independent research lab. We attack autonomous AI agents — prompt-injection chains, MCP and tool exploits, identity abuse — then publish reproducible findings and the defenses that actually hold.
$ ./secagentlabs --scope autonomous-agents [*] mapping attack surface ... [+] prompt injection (indirect) reachable [+] tool / function-call abuse reachable [+] mcp supply chain reachable [!] exfiltration path confirmed > hardening playbook ready
Then we rebuild them stronger.
Guardrails, I/O filtering, least-privilege tool design and isolation (seccomp, gVisor, WASM, microVMs) — tested against the attacks above, including their bypasses. Defenses that survive an adversary, not a checklist.
// research domains
Six fronts on agent security.
Hub-and-spoke research across the autonomous-agent attack surface — from the first injected token to the last byte that leaves the trust boundary.
ATK-01flagship
Attacking AI Agents
The attack surface of autonomous agents — prompt-injection chains, tool abuse, confused-deputy and exfiltration, with reproducible proofs of concept.
Read →MCP-02
MCP & Tool Supply Chain
Securing the Model Context Protocol and agent tools — malicious servers, tool-description poisoning, CVE teardowns and sandboxing.
Read →DEF-03
Defending & Hardening
Engineering defenses that hold — guardrails, I/O filtering, least-privilege tools and isolation (seccomp, gVisor, WASM, microVMs).
Read →RED-04
Red-Teaming & Evals
A reproducible lab for breaking agents — test harnesses, tooling (Garak, PyRIT, Promptfoo), security benchmarks and CTF-style challenges.
Read →IAM-05
Identity & Access
Identity for non-human agents — cryptographic identity, OAuth/OIDC, capability-based security, token theft and zero-trust enforcement.
Read →INT-06
Threat Intel & Teardowns
The autonomous-agent threat landscape — incident teardowns, novel attack classes and mapping to MITRE ATLAS / OWASP Agentic.
Read →// method
Evidence, or it didn't happen.
01
Reproducible
Every finding ships with a proof of concept you can run — not a screenshot, not a claim.
02
Vendor-neutral
No sponsorship in the research. We name what breaks and what holds, whoever built it.
03
Mapped to standards
Findings tie back to CVEs, papers, OWASP Agentic and MITRE ATLAS — traceable, citable.
Found a way to break an agent?
We coordinate responsible disclosure and cite reporters. Research collaboration, corrections and tips welcome — in English, Polish or German.
contact@secagentlabs.com →