The list. Method Security is on the 2026–2027 Fortune Cyber 60 in the early-growth group. Fortune’s profile lists New York City, CEO Sam Jones, 33 employees and a 2023 founding date. Fortune reports $26 million in funding from a16z and General Catalyst. That matches the company’s own announcement from November 2025: a seed led by Andreessen Horowitz and a Series A led by General Catalyst, announced together, with Blackstone Innovations Investments and others joining. Valuation and revenue are Undisclosed. Lightspeed, which co-runs the list, is not among the investors Fortune or the company name.
What it sells. Method builds AI that attacks your own systems before someone else does. Two products. Bastion sends a swarm of AI agents to map a large company’s networks, check exposures and test whether an attacker could actually get from A to B. Reaper is a workspace for professional offensive teams to plan, rehearse and run cyber operations. That second one is not a phrase you see on most startup websites.
The team. Jones spent about 11 and a half years at Palantir, much of it building its cyber and Defense Department business, after serving as an Air Force cyber operator, per a TBPN interview. Co-founders Sean Hacker (CTO) and Daniel Kelly (President) round out a founding team the company says comes from Palantir, NSA, the Air Force, CIA, CrowdStrike, Shield AI, Microsoft and AWS.
Why it is winning: it sells like a defense company. Most security startups sell to the Fortune 500 first and chase government later. Method did the reverse. In December 2025 it partnered with OpenAI to put OpenAI’s models into its platform for Defense Department and intelligence community cyber operators. In September it won a $30 million STRATFI award from the U.S. Space Force, which the company says is the maximum for that program, to test the cyber resilience of space systems. On Oct 28, the day the Cyber 60 came out, it announced FedRAMP High authorization through Palantir FedStart. FedRAMP High is the government’s strictest cloud security baseline for sensitive unclassified data, and a hard, slow climb for a 33-person company alone. Method got there through Palantir FedStart, which lets startups run on Palantir’s already-authorized infrastructure.
Notice the numbers. A $30 million government award against $26 million in disclosed venture funding. An award is not the same as booked revenue, and Method has not said how the money flows. But it tells you where this company’s weight is.
The unique technical bit: no shell, and no target it hasn’t seen. An AI that attacks is only useful if you can trust it not to go too far. Most of the industry’s answer is a second AI model that watches the first one and blocks bad moves. Method’s CTO argues that is not good enough. Citing published research, he writes that even the best “judge” models let through roughly one in ten out-of-scope actions that looked authorized.
So Method takes two things away from the AI. First, the shell. The AI never gets a raw command line. It can only use pre-built “Tools” with fixed behavior. An HTTP tool might be allowed to read a web page but not delete anything. A fuzzing tool declares up front how many requests it will send. New tools get built with a human in the loop, before the mission, not improvised live against a target.
Second, the targets. Everything the platform discovers goes into a typed map of the world it can act on: IP addresses, domains, certificates and how they connect. Method calls it an ontology. Tools can only act on things in that map. Something the system has never seen has no name, so it cannot be targeted. Rules of engagement attach to types, not lists. Write “no tools that change anything on U.S.-based IP addresses,” and a server discovered three days into an operation inherits that rule the moment it appears. The AI cannot edit those rules.
That is the real product. Attack tools are everywhere. Proof that an autonomous attacker stayed inside the lines is rare, and it is exactly what a general or a bank’s risk committee will ask for.
The CyberMerge view. Method is less a competitor to Pentera or Horizon3.ai on a bake-off sheet than a bet that cyber will be bought the way drones are bought now: by mission, with government money opening the door to commercial customers. That worked for Palantir and Anduril. It is slow, lumpy and political.
The risks are real. Thirty-three people serving the Pentagon, the Space Force and the Fortune 500 at once is a lot of surface. Government programs stall. And “we fixed AI safety with design” is a strong claim that needs outside testing, not a blog post. Watch for a named commercial customer, a production Space Force deployment, and a third party trying to push the agent out of scope.
Bottom line. Every vendor says its AI is safe. Method says its AI cannot even point at the things you told it to leave alone.
