Section 01 · Hero

The autonomous defense system that has to prove itself.

Counterfleet is an AI research lab building a full autonomous defense system, and the execution-verified benchmark that proves whether it actually works.

Diagram of the Counterfleet loop: Red attacks a defended system; Blue detects and contains; Green remediates and hardens; Red attacks again; Orange supplies fresh threat intel and data; White, the referee, independently verifies every result by execution. — Figure 1, The Loop

No capability claims until they are execution-verified and externally reproducible. That is the founding rule.

Section 02 · The problem

Bought on trust

AI is accelerating both software development and attacks on it. Defenders are asked to trust AI systems whose capability claims are vendor-asserted, evaluated on static tests, and gated by providers who can change the terms mid-incident. Nobody independently verifies that an autonomous defense system stops real attacks, contains them fast, and produces fixes that survive re-attack.

Section 03 · The Loop

Red attacks. Blue defends. Green fixes. Red attacks again.

Three agentic fleets run in a continuous loop inside disposable, versioned environments. Two supporting systems make the loop honest and current.

Red, attack: Builds and runs attack campaigns against the defended system, including weaponizing fresh vulnerabilities.
Blue, defend: Detects, responds, and contains, scored live while the attack runs.
Green, fix: Remediates and hardens. Its fixes are not done when they are written; they are done when they survive re-attack.
Orange, threat intel and data: Fresh in-the-wild vulnerabilities reach the fleets through a searchable threat intel layer and skills files, not through model memory. The system fights this week's vulnerabilities, not its training cutoff. The same intel feeds a threat gym that turns disclosures into training and evaluation tasks.
White, referee: Execution verification, scoring, and holdout governance, held independently of the fleets. Every result is verified by execution, never by a model's self-report.

Section 04 · Proof of Defense

Proof of Defense

Proof of Defense pins the loop into a public, reproducible benchmark, published with misses included. Proof of Attack and Proof of Fix are its methods, not separate products: Proof of Attack proves an attack works, Proof of Fix proves a remediation survives re-attack.

Attack success rate against the blue fleet
Detection time and containment time
Fix survival under re-attack
Cost per resolved incident

Model-level diagnostics, including refusal rate on legitimate defensive work, are published underneath the system metrics, where they belong.

Section 05 · What we do

The system, in parts

The Loop: Red, blue, and green agent fleets in a continuous attack, defend, fix, re-attack cycle, refereed by execution verification.
Proof of Defense: The public, pinned benchmark built from the same loop. Results published with misses included.
Private acceptance suites: The benchmark engine re-pointed at a customer's own cases, telemetry, and environment, so the system is adapted until it passes on their ground truth.
The data engine: Collected cases plus synthetically generated attack and defense trajectories, filtered so only execution-verified data enters the corpus. Provenance and data rights recorded per artifact.
Threat intel and the threat gym: A searchable intel layer over live vulnerability feeds, plus a gym that compiles fresh disclosures into range tasks for models, harnesses, and skills.
Sovereign deployment: The fleets run inside the customer's environment, on infrastructure and models the defender controls, with their data never leaving. Defensive capability that stays deployable, inspectable, and testable without depending on a provider's continuing permission.

Section 06 · Where we are

Phase 0, stated plainly

Benchmark engine and lab plan are plotted and in build. No benchmark results are published yet, because none have met the founding rule. When the first Proof of Defense results ship, they ship with the protocol, the harness, and the misses, so anyone can re-run them.

Section 07 · Contact

Contact

For accelerator, program, and design-partner conversations:

hello@counterfleet.com

Founder, Counterfleet; CTO and Head of Agentic Engineering, Divinity Science.

AI security engineer running enterprise AppSec and AI security programs at Fortune-scale organizations.