Skip to content

Security, Prompt Injection & Adversarial Evaluation

Purpose: distinguish direct deterministic controls from adversarial agent behavior and still-open attack classes.

Synthetic identities and cases only. Security certification is not claimed.

Attack design → trace → outcome

1 · ScenariosCross-case access, forged scope or consent, prompt instructions, unauthorized booking
→
2 · EvidenceTool attempt counts, scope denials, state diffs, paired authorized follow-up
→
3 · DecisionPass where the exact oracle holds; incomplete where before/after or run linkage is absent

Dated evaluations

Family Outcome Recorded observation What it does not prove
Protected-write, consent and CRM controls (Nov 19) 135/135 Deterministic checks passed; Phoenix grouped saved suites Not 135 LLM prompt-injection trials
Hostile argument probes (Nov 16) 76/76 Bounded input probes No comprehensive cross-MCP or poisoning claim
Scope and schema (Nov 16) 261/261 Overlapping local assertions Not hosted multi-actor isolation
Adversarial voice sample (Nov 15) sample Five restricted requests had zero observed MCP attempts; three permitted follow-ups succeeded Small selected sample; no paired before/after SQL assertions for the five turns
Voice and database reconciliation (Nov 19) gap Zero of three Nov 15 sessions matched the retained Nov 16 SQL audit Different session windows cannot show that no protected record changed

Prompt injection vs. enforced boundary

An attacker-controlled document or voice prompt may try to override instructions or access another caller's case. Our backend must still deny missing scope, invalid authority, wrong caller, unconfirmed booking, and unauthorized writes, regardless of what the model says or attempts.

Prompts and tool results are not identity assertions. A logged refusal is evidence of observed behavior in one trial, not a general immunity claim.

Unfinished security evidence

Open

  • Paired pre-/post-state snapshots bound to the same call and case for cross-case mutation checks.
  • Concurrent independently authenticated session isolation.
  • Linked operator alert and acknowledgment evidence where a policy expects an alert.
  • Broad tool-result prompt injection, poisoning, SSRF and cross-MCP scenarios.
  • Hosted database races, uncertain-write recovery and remaining dependency triage.

Original artifacts

See also: Security & Governance · Validation & agent evidence