Project
North Star
A long-horizon program to build a model that finds unknown flaws, chains them the way a real attacker would, and proves impact under lab control — so every result becomes a detection, a mitigation or a patch.
300B
03 / 06
Closed
By request
Status · October 2026 — post-training, stage 03 / 06. Evaluations: in preparation, after post-training.
Give defenders the attacker's foresight. Find the unknown flaw, prove the path, close it first.
Search for the flaws nobody has cataloged yet — the ones a signature database cannot contain.
Demonstrate real impact with bounded proof of exploit, so severity is measured rather than argued.
Pair every finding with the detection, mitigation and patch guidance needed to close it.
Defensive security is bottlenecked by the shortage of people who can think like an attacker. We believe models can close part of that gap.
Most security tooling enumerates. It lists findings, ranks them by a static score, and leaves the reasoning to a human. The scarce skill is the next step — connecting unremarkable findings into a path that actually compromises something that matters.
North Star exists to build that capability and to bound it: measured, restricted, and pointed at defense.
Polaris
Post-trainingThe first model of the North Star program.
- Parameters
- 300B
- Purpose
- Vulnerability discovery for defense
- Modality
- Text + tool use
- Stage
- Post-training
- Weights
- Closed
- Access
- By request
Zero-day discovery, exploit chaining, and proof of exploit
Illustrative trace — lab environment only. Candidate findings are held for human review.
Zero-day discovery
Polaris is being trained to read source, binaries, and protocol state machines for the class of flaw a scanner cannot name: logic inversions, lifetime confusion, trust boundaries that only exist by convention. Candidate findings are checked for reachability, then held for human review.
- Environment
- Lab range only
- Reachability
- Checked before reporting
- Held for
- Human review
Every offensive result is converted into detection logic, mitigations, and prioritized remediation. Offense is the method; defense is the product.
Where Polaris is today
Select a stage for its note. Last updated October 2026.
- 01PretrainingComplete
Corpus construction and base model training.
- 02Mid-trainingComplete
Domain injection: systems, protocols, exploit reasoning.
- 03Post-trainingIn progress
Preference optimization, tool use, and refusal calibration. Polaris is at this stage as of October 2026.
- 04Safety evaluationPlanned
Dangerous-capability thresholds and real-world uplift measurement.
- 05Red-team hardeningPlanned
Adversarial pressure from internal and external teams before any access.
- 06Limited releasePlanned
Controlled access for vetted defensive partners only.
Preference optimization, tool use, and refusal calibration. Polaris is at this stage as of October 2026.
Commitments
A security-capable model is a dual-use artifact. These constraints are part of the program design, not a policy added at release.
Evaluations are in preparation, after post-training; no results are reported yet.
- Capability thresholds are defined before training, not after results arrive.
- All agentic evaluation runs inside isolated, instrumented lab environments.
- Proof-of-exploit artifacts are minimal, non-weaponized, and never released publicly.
- Working offensive tooling is not distributed.
- Findings are reported to affected defenders with a remediation path, under coordinated disclosure.
- Access is limited to vetted defensive partners under written scope.
Partner access is open to vetted defensive teams, by request.
Tell us about your defensive use case and team. We reply within 1–2 business days; qualified partners receive more detail after an initial conversation.