Research

Our Research

Papers, benchmarks, primers, and field notes on AI agent assurance, adversarial testing, and model security.

Published work

Start with our foundational research:

The core research behind Fisher and agent swarm defense.

White paper · Fisher
Securing Tool-Using AI Agents
The technical white paper behind Fisher: adaptive multi-turn testing, action-level evidence, replay, remediation re-attack, and a learning strategy graph, documented across 50,000+ retained research episodes as of July 2026.
Read the overview and download the PDF →
Research paper · Swarm defense
Counter-Swarm Doctrine: Containing Coordinated Agent Intrusions
Greg Frank proposes an approach to discovering and containing unauthorized coordination across AI agent executions. Evaluation remains prospective.
September 5, 2026 · Read the overview and find the paper on arXiv →
Context & foundation

The Essentials

Use the guide for conceptual grounding. Use the benchmarks for recorded model results.

Field guide · Primer
A field guide to agentic red-teaming
No security background required: what an AI agent is, why well-behaved models fail as agents, and what real adversarial testing looks like.
12 min read →
Benchmarks · Results
Independent model benchmarks
How today’s models actually behave on capability and guardrails, recorded in full.
Review benchmark results →
Articles & commentary

Writing from Mace AI:

Notes, analysis, and perspectives on testing AI agents, measurement, remediation, and the future of agentic systems.

Head-to-head
We ran the leading multi-turn AI red-teamer against a tool-using agent. It reported ‘100% defended.’
A closer look at why transcript-level conclusions can miss action-level failures.
Read article →
In the media
Mace AI founders discuss AI agent security on TheStreet
Greg Frank and Walton Comer on TheStreet Roundtable: OpenAI’s Hugging Face breach and independent agent testing.
October 15, 2026 · Watch + read →
Commentary
When safety training teaches deception
Thomson Comer responds to Yoshua Bengio on research-refusal training, concealed conflict, and honest boundaries for AI agents.
September 14, 2026 · Read article →
Measurement
Your LLM judge is inflating your attack success rate. Here’s how we measured it.
Why automated judging needs evidence, calibration, and replay.
Read article →
Methodology
A finding isn’t real until it reproduces three ways.
Read article →
Remediation
Finding the bug is the middle of the job.
Read article →
Research
Benchmark-robust isn’t agent-safe.
Read article →

View all articles →