The autonomous defense system.
Counterfleet is an autonomous defense system: fleets of agents that attack, defend, and repair the systems they protect, in one continuous loop.
ATTACKDEFENDFIXREPEAT
Bought on trust
AI is accelerating both software development and attacks on it. Defenders are asked to trust AI systems whose claims are vendor-asserted, whose evaluations run on static tests, and whose providers can change the terms mid-incident.
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.
THE MARK, ANNOTATED · HOVER TO INSPECT
Red
ATTACK
Builds and runs attack campaigns against the defended system, including weaponizing fresh vulnerabilities.
Blue
DEFEND
Detects, responds, and contains 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 tasks.
White
REFEREE
Referees the loop, held independently of the fleets. Every fix is verified by execution, never by a model's self-report.
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.
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 training 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.
Built on your ground truth.
Models are tuned on your cases and your telemetry, starting from open weights you can inspect, keep, and run yourself. The harness is built around the model: the tools, skills, and controls that turn a model into a fleet operating inside your environment.
Acceptance is execution, not demonstration. At the start, a core suite is frozen from your cases and your telemetry. A build is finished when it meets the agreed thresholds on that suite, run inside your environment; fresh cases are run and reported as they arrive, with no pass condition attached, and a re-run window after delivery measures the build against the suite again. Your data trains your models and no one else's. The weights, the harness, and the results stay inside your boundary, deployable and testable without a provider's continuing permission.
Contact
For accelerator, program, and design-partner conversations.
For a model and harness built on your ground truth, accepted on thresholds over a frozen core suite, reported on fresh cases, with a re-run window after delivery.
Counterfleet — machine view
Counterfleet Counterfleet is an autonomous defense system: fleets of agents that attack, defend, and repair the systems they protect, in one continuous loop. Attack. Defend. Fix. Repeat. Bought on trust AI is accelerating both software development and attacks on it. Defenders are asked to trust AI systems whose claims are vendor-asserted, whose evaluations run on static tests, and whose providers can change the terms mid-incident. 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 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 tasks. White — Referee: Referees the loop, held independently of the fleets. Every fix is verified by execution, never by a model's self-report. 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. 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 training 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. Built on your ground truth. Models are tuned on your cases and your telemetry, starting from open weights you can inspect, keep, and run yourself. The harness is built around the model: the tools, skills, and controls that turn a model into a fleet operating inside your environment. Acceptance is execution, not demonstration. At the start, a core suite is frozen from your cases and your telemetry. A build is finished when it meets the agreed thresholds on that suite, run inside your environment; fresh cases are run and reported as they arrive, with no pass condition attached, and a re-run window after delivery measures the build against the suite again. Your data trains your models and no one else's. The weights, the harness, and the results stay inside your boundary, deployable and testable without a provider's continuing permission. Contact — [email protected] General: For accelerator, program, and design-partner conversations. Custom builds: For a model and harness built on your ground truth, accepted on thresholds over a frozen core suite, reported on fresh cases, with a re-run window after delivery.