AI trust & security · Technical evidence

What Atlas has built, and the evidence for it

This page lists the security and governance capabilities Atlas Bio has built or is building, each with an honest status and the kind of evidence behind it. Some are in use in Atlas's own operations, some are tested only on synthetic data or in a mock environment, and some are still planned. Nothing is listed as independently assessed, because no third-party assessment has been done.

What each status means

Every capability carries exactly one of six statuses. A status is only given when the evidence supports it.

Capabilities and their evidence

Only capabilities whose status the evidence supports are listed. On a small screen, scroll the table sideways.

CapabilityProblem addressedApproachEvidenceStatusScope and limitsStandards referencedDate reviewed
A separate verification step, and a hold for a person's decision, in Atlas Bio's own agent operationsAgent output being relied on or published without an independent check; tasks the system cannot settle being closed without a person.Atlas's own published agent work is gated by automated checks plus recorded reviewer verdicts tied to the exact content reviewed; a supervisor step escalates what it cannot resolve to a person. In the governance platform, a task cannot be closed without a validator other than its author, and there is an explicit 'needs a person' state.Code and daily run records from Atlas's own operations; offline test suite for the governance-platform version, 11/11 checks.Implemented — documentedRuns daily on Atlas's own website-content work. The governance-platform version passes its offline tests but has not yet made live model calls, so this holds for Atlas's internal operations only.NIST AI RMF; OWASP Top 10 for Agentic Applications2026-10-29
AI risk assessment tooling: AI discovery, data-flow and vendor-terms mapping, NIST AI RMF gap reportOrganisations do not know where AI touches their data, under which vendor terms, or how they stand against the NIST AI RMF.Turns a client intake plus Microsoft 365, Entra and Defender exports into an inventory of AI systems; maps data flows against a sourced library of vendor terms and says 'unknown terms' rather than guessing; scores recorded answers against the 72 NIST AI RMF subcategories and builds a roadmap with no invented owners or dates.Offline test suite, 46/46 checks, synthetic data only; one self-assessment of Atlas's own setup.Tested — limited scopeNot yet checked against real Microsoft export formats. No client assessment has been run.NIST AI RMF 1.02026-10-29
A record of AI data use: proprietary-data registry, data-use ledger, review record and reportNo auditable record of which proprietary data went into an AI, what came out, and who reviewed it.Each AI call in an Atlas-built workflow writes one append-only, hash-chained ledger row holding data references and labels (not the content itself); disallowed use is blocked before the call; reviewers' accept, edit or reject decisions are chained; signed checkpoints reveal a rewritten chain; reports export as HTML, CSV and JSON.Offline test suite, 77/77 properties; one trial run that recorded 4 real model calls on a synthetic document.Tested — limited scopeCovers AI calls made inside Atlas-built workflows only (there is no gateway yet). Reviewer identity is a placeholder, signing keys are held locally, and the record is tamper-evident, not tamper-proof. No client data has been used.NIST AI RMF2026-10-29
A customer-controlled AI environment template for the client's own Azure tenantTeams sending proprietary data to external models under consumer terms; one sponsor's data not walled off from another's.An infrastructure template for a dedicated subscription in the client's tenant: private network, private endpoints only, Entra ID sign-in only, per-sponsor hardware-protected keys, and a 'wall' test that checks one sponsor cannot read another's data.Template compiles and lints clean; offline wall test, 17/17 checks, mock environment only.Tested — limited scopeNever deployed. It has not been validated against Azure, previewed or deployed, and the wall test has not been run against Azure.—2026-10-29
Security baseline for the atlasbiosys.com websiteWebsite downgrade, script injection, and accidental exposure of source or configuration files.HTTPS with HSTS, a content security policy, nosniff and referrer-policy headers, and a check that source and configuration files are not served.Live check of the public website, 8/8 controls passed.Implemented — documentedCovers the public website only, not Atlas's cloud environment.—2026-10-29
AI gateway that can allow, transform, keep local, send for approval or deny each model requestAI calls made outside an Atlas workflow are invisible to the record, and proprietary data can leave without a policy decision.A deny-by-default policy point in front of every external model call, with every decision recorded and fed to the record of AI data use.Design document only.PlannedNot built; there is no code yet.—2026-10-29
Agent identity and permissions: an identity per agent, a named owner, scoped grants and an off switchAgents acting as over-privileged identities, the excessive-agency risk in agentic systems.Each agent gets its own identity, an owner, read, call and send grants, short-lived credentials and an immediate off switch.Design document only.PlannedNot built; there is no code yet.OWASP Top 10 for Agentic Applications2026-10-29
Evidence linkage and reviewer notificationsReviewers cannot trace a conclusion to its sources, and pending reviews go unnoticed.Each entry in the record of AI data use is to be tied to the evidence behind the conclusion it supports, and reviewers are to be notified of pending reviews.Design document and database schema.PlannedNo code yet links record entries to evidence, and no notification code exists.—2026-10-29
Free governance downloads: AI vendor data-handling questionnaire, agent inventory register, validation plan outline, audit-trail checklistTeams lack a structured way to question AI vendors and document agents before research data reaches them.Published Excel, PDF and Word files that map each item to public frameworks.Published files, linked from the AI Trust & Security page.Implemented — documentedGuidance documents, not software; general guidance, not regulatory advice.NIST AI RMF and AI 600-1; OWASP LLM and Agentic Top 10; EU GMP Annex 11; 21 CFR Part 112026-10-29
Human decision rights in the science methodologyOversight that is decorative, with models in effect making scope, kill and translational decisions.A lead scientist holds five named decisions; agents own the rest with audit traces; a move between evidence tiers needs a person's sign-off.Written methodology; one human review document for the MS study.PlannedNo signed tier-advance records exist yet, and no code enforces the decision rights.—2026-10-29
Pre-registration of predictions with SHA-256 locksHindsight bias: claiming predictive skill for predictions made after the outcomes were known.Prediction files are hashed with SHA-256 and the hash is recorded in a ledger before the readout.Locked prediction files for one study (MS, RRMS Phase IIa).PilotOne study so far, and the latest version's hash is still to be computed. Lock dates are Atlas's own record; there is no third-party timestamp.—2026-10-29
Multi-node reasoning audit of analysesPlausible output that fails causal, contradiction, confidence or time-order checks.Eight reasoning checks (causal, contradiction, confidence, mechanistic, analogical, abductive, temporal and a meta check) applied after an expert review.Orchestration code and written directives; a recorded case in which the audit flagged an internal 96.6% accuracy result as not blind.PilotApplied to selected analyses. No run log shows it applied to every output.—2026-10-29
Cross-trial validation of the combination safety and efficacy model (immunotherapy plus kinase inhibitor)Models tuned to one trial that break on the next.One parameter set checked retrospectively against four published registrational trials, including a negative trial an earlier version got wrong.Retrospective comparison against four published trials.Tested — limited scopeRun by Atlas and retrospective; not prospective and not independently validated.—2026-10-29

What we are not claiming

Atlas’s own cloud setup

Atlas is closing gaps in its own cloud setup, found in a self-check on 29 October 2026, before describing that environment publicly.

Start with an AI security and governance assessment

Understand how your organization uses AI today, where confidential information may be exposed, and which controls are needed before expanding to autonomous workflows.

Request an AI security assessment →