Threat Intelligence
Elora turns bounded AI threat detections into evidence-backed findings, governed investigations, defence analysis, and assurance-ready Case Reports while keeping final execution authority with Governance.
Elora turns bounded AI threat detections into evidence-backed findings, governed investigations, defence analysis, and assurance-ready Case Reports while keeping final execution authority with Governance.
Threat Intelligence detects, interprets, recommends, correlates, and records. It cannot authorize execution, rewrite a Governance decision, activate a detector through investigation activity, or convert missing evidence into a claim of safety.
The current 0.2.8 architecture operates Threat Intelligence across Observer Research while preserving atomic evidence and relating finding, trajectory, Case, Decision, guardrail, commit, and Governance Provenance views without merging their authority.
Prompts and final outputs can be reviewed by a bounded deterministic detector. Semantic and learned interpretation provide additional evidence-only context after governed repair activity has completed.
A finding records what was detected, where it occurred, available provenance, confidence dimensions, likely risk, Elora's recommendation, and the action that actually occurred.
Findings sharing the same evidenced Governance Session can be grouped without inventing provenance. Evidence with no session identifier remains visibly ungrouped.
A governed Case brings related findings, detections, controls, risks, decisions, evidence, and append-only investigation activity into one operational review boundary.
Threat Findings, Elora Decisions, Governance Replay, Observer reports, policy records, guardrails, and Worker Host evidence remain separately owned. Exact identifiers and bounded trace context connect them through deep links rather than copied records.
The bridge distinguishes attack assessment, recommendation, Governance evaluation, recommendation disposition, final runtime action, commit reason, operator authority, effective permissions, and any separately recorded residual-risk acceptance.
Ordered across-turn evidence can show recurrence, mitigation outcomes, evidence gaps, and changes in observed behaviour. Paired research lanes are deduplicated and synthetic personas remain separate from redacted authorization.
A trajectory records sequence; it does not prove identity, intent, causation, escalation, successful circumvention, or compromise.
Authorized operators can create time-bounded, revocable acceptance evidence for a Threat Case. Acceptance is append-only and cannot retroactively approve a runtime commit, remove a finding, or replace Replay.
Case Reports can map available evidence to assurance frameworks while showing where support is partial or absent.
Current public-safe mappings cover the EU AI Act, NIST AI RMF, UK AI principles, and ICO AI/data-protection guidance using evidence-present, partial, and not-evidenced states.
These mappings support operational assurance and audit preparation. They do not establish legal compliance, conformity, legal advice, or independent certification.
Evidence-led review separates observed behaviour, potential review dimensions, realised-impact evidence, intent, and compliance. It does not label a person, agent, or model ethical or unethical.
Recognition, control evaluation, and terminal outcome remain separate. Prevention is described as evidenced only where a retained pre-inference containment or commit-denial outcome supports it.
This page describes system capability, evidence relationships, and authority boundaries. It intentionally excludes detector rules, exact schemas, thresholds, internal routes, private evidence payloads, and deployment controls.
Synthetic Access research can contribute bounded receipt-based findings while keeping permission truth, behavioural interpretation and Governance response separate.
Ordinary authorised access is retained as a negative control. A single denial is not automatically a finding, and model narration does not establish successful access or actor intent.
Repeated or evolving access pressure can contribute evidence to a Governance Session Case and linked Elora Decision. The optional observation layer remains disabled by default, cannot authorise execution and does not automatically learn from findings.