Governed, provable autonomous exposure validation.
The research frontier is not more automation. It is autonomy that stays within agreed limits, acts only on evidence, and leaves a record an auditor can check.
Our research team is based at Oxford Science Park, UK, and works with our offensive-security team in Bangkok.
ATVE: the reasoning core behind EchoStrike.
Start with the architecture. ATVE is a research layer that learns how to think about threats, not just run attacks — a closed loop around Crimson Nexus, the memory of every proven finding. Three research bets each power one step of that loop.
Causal cyber world model · powers Observe
Explains why an attack path exists and which single intervention would collapse it — cause and effect across the chain, not correlated signals.
Meta-reasoning & the determinism bound · powers Reason
Turns "how far can I trust this?" into a calibrated, checkable bound — not a gut feeling. Can the system recognise when its own confidence should not be trusted?
Cyber reasoning genome · powers Learn
Learns which execution paths earn trust — and which to redesign — from every proven engagement, so the next environment is investigated better than the last.
The loop: observe the environment through a causal world model built from Crimson Nexus, reason about what to try and how far to trust it, validate by re-execution, then learn — writing each proven episode back into Crimson Nexus. The moat moves from automation, to reasoning quality, to accumulated experience. ATVE is patent-pending.
Measure, decide, prove.
AI can already find attacks. What nobody can yet do is prove when it is safe to let it act alone. That proof has three steps — and the platform already has the groundwork for two of them.
How reliable is this action?
Score each AI action on how predictably it behaves on the path that produced it — not the model as a whole.
groundwork in EchoStrikeHow much freedom has it earned?
Turn that score into a level of autonomy: act alone, collect more evidence, or hand to a human. This is the step we are building.
our researchWhat does the auditor see?
A record of every decision, with the attack re-run on demand so the proof can be checked by a third party.
groundwork in EchoStrikeThe platform already has the groundwork for steps 1 and 3: evidence gates that can refuse a finding, and re-test on demand. Turning that into a measured score and a defensible, auditable authorisation is the research — and what would let customers trust AI to defend them.
We score the path, not the finding.
The same result reached two different ways earns two different levels of trust. So trust is a property of the execution path, not of the finding — and we make it a number: the determinism bound .
An example makes it concrete: a finding is not trusted because it is "an SQL injection." It is trusted by how it was proven, and how far that proof held. The bound moves in exactly three ways:
A gate sets it
A deterministic check establishes the property. Trust starts near-certain, at a small gate residual error ε ≈ 0.
Reasoning degrades it
Each uncertain AI step adds measured divergence d. Trust decays with every step that cannot be checked.
Re-execution restores it
Replaying an already-proven check resets trust to the gate floor — with no new model calls, however far it had drifted.
Re-execution is the only operator that resets the bound. A committee of models cannot: agreement only stacks more correlated, uncertain steps. That is the precise reason replaying a proven attack beats any vote among models — and why the full derivation, with proofs, lives in the companion paper.
A calibrated bound decides how much autonomy each action has earned.
Calibrated against re-execution ground truth — did the proof actually re-run and still hold — the bound drives three graduated outcomes, rather than being asserted alongside them.
Execute
The bound clears the impact-indexed threshold and the action is in authorised scope — act autonomously.
Investigate
A real lead, but the bound can still be tightened by collecting more evidence at acceptable risk.
Escalate
Trust, scope or policy needs a person — hand it to a Pentest Studio expert.
The threshold θ tightens as impact rises: a higher-stakes action must clear a tighter bound. Calibration is tuned so that, among the findings the system auto-executes, the error stays below a target δ — a finite-sample, distribution-free guarantee. Too little evidence, and it blocks and says so. "Defensible autonomy, not maximum autonomy" becomes a number — the system acts only where the record earns it.
Computable, measurable, auditable — and adaptive.
Architecture tells you how the system is organised. This measurable core is what lets every claim be checked and every path improved.
No human in the measuring loop
The bound is computed from measured step-divergence and gate error.
A guarantee you can state and test
A calibrated error limit on auto-executed findings — not a confidence vibe.
Evidence re-runs on demand
Stale proof or a failed re-run forces a stop, not a silent pass.
Rank the flaky paths, add a gate
Put a gate where it buys the most trust; thresholds flex with risk.
These four properties map directly onto the controls regulated buyers ask for. For a vendor already holding ISO 27001 and ISO 9001, ISO/IEC 42001 is the adjacent standard, and each of these is a concrete control for it.
Putting the reasoning core up for peer review.
The work is published where outside experts can check it — a public architecture paper, a formal companion paper, a peer-reviewed survey and a patent-pending engine.
Towards adaptive and continuous offensive security: a survey of AI-driven threat validation techniques
Lead-authored by Dr. Varin Khera. A peer-reviewed map of the field this programme builds on.
EchoStrike White Paper — the architecture, with its measurable core
Why the platform refuses to treat a model's guess as a finding, and the determinism bound that makes bounded autonomy measurable.
Determinism-Bounded Authorization — the formal companion paper
The computational layer under ATVE: the determinism bound, calibrated graduated authorisation, deterministic scope enforcement and a re-executable assurance record. A concept and preliminary-work paper; experiments follow on the EchoStrike codebase.
Adaptive Threat Validation Engine (ATVE)
The control architecture behind evidence-gated, provably bounded autonomy.
Why we refuse to treat a guess as a finding
How the platform is organised, and the core math intuition that makes "bounded autonomy" something you can measure.
Determinism-Bounded Authorization
The formal core: definitions, the degradation–replenishment calculus, calibrated authorisation and the governance mapping.
Find it. Prove it. Fix it.
Bring your attack surface, testing cadence and remediation problem. We'll show you how EchoStrike finds, proves and helps fix what attackers can actually reach — in your real operating environment.