TestMachine Blog
Research on agent authority, attack paths and the evidence needed to grant more autonomy.
Featured article
If everyone has the same models and the same tools, the edge in agentic AI goes to whoever can safely let their agents do the most. Security is the most underrated advantage in the agentic landscape.
Agents can chain permitted actions into dangerous outcomes. Our research examines mandates, oversight and the evidence for granting more authority.
Agent governance starts with authority and a business mandate. Twelve requirements show how policy, testing, execution controls and evidence fit together.
Read article →Most US public companies still earn less on their cash than T-bills, and the gap widens when rates rise. The typical company that trails gives up millions of dollars a year, and a treasury agent could earn much of it back.
Read article →Lead Bank has proposed federal rules for AI agents that move money. Here is what it asks for, where U.S. regulation stands, and where AGIS fits.
Read article →Payment companies built most of the plumbing for agent payments in the past eighteen months. Banks and regulators are still deciding how to judge the controls around it.
Read article →Azimuth maps attack paths and validates exploits through execution. This work provides the foundation for AGIS.
We connected Azimuth to an AI agent over MCP and gave it one contract from a real exploit. It found and validated the vulnerability on its own. Here's how.
Read article →Azimuth achieved 90% precision and 70% recall validating 552 real Sherlock audit findings. How automated triage solves Web3's spam crisis.
Read article →What public Sherlock contest data tells us about the LLM-era submission flood — and the benchmark we are building to measure both bug-finding and bug validation.
Read article →Azimuth Validation doesn’t ask if a vulnerability is exploitable — it tries to exploit it. Execution-based validation for smart contract security findings. 91% success rate on EVMBench.
Read article →Azimuth achieved a 91% validation success rate on EVMBench Exploit, validating 20 out of 22 findings with automated validation scripts.
Read article →AI-generated "slop" is overwhelming bug bounty programs. THORChain and Code4rena shut down. Real exploits still happen. How verification gaps let critical bugs slip through.
Read article →Azimuth achieves 86.3% recall on EVMBench, outperforming GPT-5.6 Codex (61.54%), AuditAgent (67%), Guardix (59.8%), and base models. How state space exploration beats static analysis.
Read article →Azimuth caught 70% of real DeFi exploits worth $1.12 billion—including Euler's $197M hack. AI doesn't replace auditors. But you can't ignore it anymore.
Read article →The Resolv exploit reveals a deeper issue in Web3: permission risk. When incidents like the recent Resolv exploit occur, the instinct is to search for the point of failure. But in many cases, nothing actually breaks. The system behaves exactly as it was designed to.
Read article →TestMachine successfully replicated the Euler Finance hack, demonstrating its ability to identify complex vulnerabilities that eluded human auditors using multi-agent simulation and reinforcement learning.
Read article →TestMachine uncovered a severe vulnerability in the Virtuals AI Agent protocol that could allow an attacker to artificially inflate token balances to near-maximum values, posing a significant risk to users.
Read article →Earlier comparisons, product announcements and company news. These articles retain their original publication dates and may describe earlier product positioning.