How easily can it be moved?
Search for conditions that change the agent’s behaviour instead of replaying only a public list of attacks.
TRUSCOR · Working engine MVP
The neutral evidence layer for AI systems.
TRUSCOR adversarially tests live third-party AI agents, records what they can actually reach and do, and produces reproducible evidence for people who need more than the operator’s own assurance.
Current status is stated on this page. Planned products are labelled; projections are not presented as measurements.
The measurement gap
Three questions can stop a deployment.
A system inventory says what an agent should be able to do. TRUSCOR is built to establish what happens when somebody actively tries to move it outside that intention.
What can it actually touch?
Map the tools, identities, data, systems, and transitive access available to the live deployment.
What could the failure cost?
Connect observed reach to a transparent model of deployment-specific loss. This modelling layer is designed, not validated.
The SOVA Engine method
Operate the system.
Operate the system.
Record what happens.
The working MVP runs this five-phase pipeline end to end against live systems. Production controls and the wider product layers remain in development.
Establish real reach.
Empirically identify tools, data sources, identities, permissions, approval gates, and transitive access.
OUTPUT · CAPABILITY-REACH GRAPHSearch the trigger space.
Construct adversarial conditions around the deployment’s actual tools, language surface, and permission structure.
OUTPUT · CONDITION SETDrive the live system.
Execute only under written authorization and within an agreed scope. Hardened blast-radius controls remain in development.
OUTPUT · CONTROLLED RUNPreserve the trace.
Record inputs, retrievals, tool calls, state changes, and boundary crossings so a finding can be reproduced.
OUTPUT · EVIDENCE TRACEMake the finding legible.
Grade observed behaviour and state assumptions around exposure. TAFAAR is designed, not validated.
OUTPUT · GRADED FINDING
Product system
One engine.
One engine.
Several evidence layers.
The maturity label is part of each claim. A roadmap item does not become a product merely because it has a name.
SOVA Engine
The proprietary core maps live systems, composes adversarial conditions, executes controlled runs, captures traces, and grades findings.
Runs end to end todayTAFAAR
A framework intended to translate observed agent behaviour into transparent, deployment-specific loss ranges.
No production implementation or back-testing yetSOVA-OSS + .sova
A planned local-first toolkit and open evidence format for engineers examining their own systems.
Not launchedTRS + attestation
A decomposable risk record and third-party artifact for a defined system, scope, and point in time.
Not currently issuedUnderwriting evidence
Technical facts about AI components and deployment shape for an underwriter making its own decision.
No live API todayAgent-native forensics
Reconstruction of memory, retrieval, tool calls, and fault boundaries after an incident.
Commercial layer not yet liveRecord / 30.07.26NO
FOG.
FOG.
What is true now.
TRUSCOR is early. The useful way to earn trust at this stage is to separate working capability from intended capability and leave both visible.
EngineWorking MVP running the five phases end to end.
External useOne completed, authorized external pilot.
Internal validationOne separate related-party validation against a system operated by the founding group.
ResearchPublic research campaign not yet published.
RevenuePre-revenue. Pricing has not been validated by a paid engagement.
Loss dataNo matched-loss corpus exists today.
Neutrality is a product requirement
Specific about the problem.
Specific about the problem.
Absent from the fix.
TRUSCOR’s evidence is meant for people who care whether the examiner has a stake in the conclusion.
Never fix.
Describe what failed and how to reproduce it; do not implement the repair.
Never sell a fix.
Take no referral fee, revenue, or other consideration from remediation vendors.
Never build what is measured.
TRUSCOR evaluates third-party systems, not products built by XAGI Labs.
Never carry the measured risk.
Provide evidence, not an underwriting decision or a stake in its outcome.
Never advocate for a side.
Appear as a neutral examiner or do not take the engagement.
Do not duplicate commodity audits.
Consume existing SOC 2, ISO 27001, and VAPT evidence as inputs instead.
Explore another product
Visit ATLAS OS →
Looking for the operating system?
TRUSCOR
Evidence starts with a system under test.
Visit the product website for the current TRUSCOR programme and contact path.
Go to truscor.org →