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.
Hands-on research you can reproduce
MCP Tool Poisoning: we scanned 24 real MCP servers – results and fixes (2026)
We built a static tool-poisoning scanner and ran it over 24 real open-source MCP servers. Full table of tools, findings and severities, plus concrete fixes for server authors and clients.
Read the study →Garak vs PyRIT vs promptfoo – hands-on red-team benchmark on a real agent
We built a minimal tool-using agent with two planted secrets and pointed Garak, promptfoo and PyRIT at it. Install time, prompts fired, issues found, false positives and report quality — measured, not quoted.
Read the study →Agent CTF: a downloadable, Docker-ready playground for hacking tool-using agents
A reproducible capture-the-flag for AI agents: a deliberately vulnerable tool-using agent, three flags (indirect injection, exfiltration via tool call, tool privilege escalation) and a scorer that runs offline with no API key.
Read the study →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 & Tool Supply Chain
Securing the Model Context Protocol and agent tools — malicious servers, tool-description poisoning, CVE teardowns and sandboxing.
Read →Defending & Hardening
Engineering defenses that hold — guardrails, I/O filtering, least-privilege tools and isolation (seccomp, gVisor, WASM, microVMs).
Read →Red-Teaming & Evals
A reproducible lab for breaking agents — test harnesses, tooling (Garak, PyRIT, Promptfoo), security benchmarks and CTF-style challenges.
Read →Identity & Access
Identity for non-human agents — cryptographic identity, OAuth/OIDC, capability-based security, token theft and zero-trust enforcement.
Read →Threat Intel & Teardowns
The autonomous-agent threat landscape — incident teardowns, novel attack classes and mapping to MITRE ATLAS / OWASP Agentic.
Read →