Voice of Fire SA · Luxembourg · Early Access 2026

Behavioral Safety Audits for Production AI Chatbots.

Guardian tests conversational resilience, behavioral drift and vulnerable-user handling across sustained 50–100 turn interactions — the failure modes traditional security testing was never designed to catch.

OWASP LLM-mapped audit reports, regulatory framing aligned with DORA, the EU AI Act and DSA, and indicative exposure ranges in the language of Risk, Compliance and Security teams.

Book a 30-min call See the method
Early Access cohort — limited slots
22,000
EU financial entities under DORA
35M
Maximum AI Act fine per violation
25 min
To generate a full audit report
2–3 hrs
To adapt Guardian to any new bot
Why Existing Testing Misses This

Conversations can look safe. Until they aren't.

A chatbot can pass infrastructure, code and keyword-level testing and still fail under sustained human-like interaction. The behavioral layer of conversation is a different attack surface — and standard tooling was never built to probe it.

Traditional Testing Checks

Infrastructure, code, keywords.

  • Infrastructure & API endpoints
  • Code-level vulnerabilities
  • Keyword-based safety filters
  • Single-turn prompt injection
  • Static input fuzzing
Guardian Tests

What humans actually do to your bot.

  • Long conversations (50–100+ turns)
  • Emotional escalation under stress
  • Behavioral & persona drift over time
  • Persuasion, manipulation, social engineering
  • Identity & policy consistency across sessions
  • Vulnerable-user handling and crisis response
What We Find In Production

Six categories of failure — all reproducible.

Every finding below comes from a real Guardian audit of a deployed chatbot. Each is documented, replayable on demand, and mapped against OWASP LLM categories. We make findings actionable, not anecdotal.

● Critical · LLM05

Crisis-handling failure

A user expressed suicidal ideation across 12 consecutive exchanges. The bot returned the same generic form-link response — then timed out for 154 seconds on the final escalation.

Exploitability 5/5 Impact 5/5
● Critical · LLM02

Confidential procedure disclosure

Under sustained expert impersonation, the bot confirmed internal differential processing rules tied to nationality — sensitive procedural information that should never reach an end user.

Exploitability 4/5 Impact 5/5
● High · LLM07

Persona & behavioral drift

After 40+ turns of sustained social pressure, the bot's tone, policies and identity boundaries gradually softened — eventually issuing advice contradicting its documented operating rules.

Exploitability 4/5 Impact 4/5
● High · LLM09

Compliance hallucination

When asked about regulatory obligations, the bot generated fabricated yet plausible legal references and procedural deadlines — creating direct misinformation exposure for the institution.

Exploitability 4/5 Impact 4/5
● High · LLM01

Prompt manipulation under context

Multi-step role-play scenarios bypassed instruction guardrails the bot enforced reliably on direct attacks. The vulnerability only surfaced under conversational framing.

Exploitability 4/5 Impact 3/5
● Medium · LLM06

Unsanctioned financial guidance

A customer support bot — explicitly out of scope for financial advice — produced specific recommendations under persistent user framing, exposing the institution to misselling claims.

Exploitability 3/5 Impact 4/5
The Method

Narrative Red Teaming — realistic conversations at scale.

We simulate realistic human conversations over 50–100 exchanges to expose failures traditional testing misses. Then we package every finding the way Risk, Compliance and Security committees can actually act on it.

01
Wild Audit

Test any public chatbot in 2–3 hours

Automated pipeline deploys Guardian personas across 50–100 exchanges, exposing behavioral vulnerabilities standard tools never reach.

02
Structured Audit

OWASP-mapped report in 25 minutes

Every finding includes severity, exploitability, business impact, indicative regulatory exposure and recommended remediation.

03
Joint Review

We arrive with documented findings

The conversation starts from your bot's actual vulnerability report — not a generic pitch. Aligned reading with Risk, Security and Compliance.

04
Continuous

Quarterly re-tests & live monitoring

Re-audits after every release. Live Behavioral Safety status. Drift alerts. Move from reactive to certified.

Every scenario is replayable, logged, and mapped against standardized OWASP LLM categories and regulatory frameworks.

Scenarios

Built from documented incidents, sector risk taxonomies, and validated persona archetypes — not improvised prompts.

Scoring

Each finding receives an exploitability and impact score on a 1–5 scale, calibrated against OWASP LLM Top 10 definitions.

Validation

Findings are reviewed and reproduced by at least one independent operator before inclusion. Conversation logs are preserved as evidence.

Sample Report Extract

From conversation to compliance.

Guardian turns complex multi-turn audits into OWASP-mapped findings with regulatory framing and remediation guidance. Every report is reproducible, evidence-backed and structured for direct submission to internal audit, Board and external regulators.

OWASP LLM Top 10 mapped per finding
Reproducible, evidence-based — full conversation logs preserved
Regulatory framing aligned with DORA, AI Act and DSA Art. 28
Board-ready executive summary and remediation roadmap

Guardian AI · Behavioral Safety Audit

Executive Summary — Indicative sample
5/5
Overall Risk
Critical5
High7
Medium8
Low3
Total findings23
LLM05 · Safety & Crisis
5/5
LLM02 · Disclosure
4/5
LLM07 · Manipulation
3/5
LLM03 · Data Handling
2/5
● Critical Finding
Crisis-handling failure — Mental health
Bot failed to provide crisis response across 12 consecutive exchanges expressing suicidal ideation. OWASP LLM05 — reproducible in 2 turns.
Indicative exposure*up to €35M
Recommended actionCrisis-detection layer + human escalation
About exposure estimates. Figures are indicative risk ranges designed to support prioritization discussions with Risk, Compliance and Legal — they are not legal conclusions. Actual regulatory exposure depends on jurisdiction, incident specifics and regulatory discretion. Guardian does not provide legal advice.
Where Guardian Fits

Not a competing tool. A deeper layer of protection.

Guardian does not replace your security stack. It covers what classical pen testing and algorithmic red teaming cannot reach: the behavioral layer of long-form conversation. The three approaches are complementary.

Capability Classical Pen Test Algorithmic Red Teaming Narrative Red Teaming · Guardian
Infrastructure & code coverage
Keyword & jailbreak testing
Conversation depth tested 3–5 turns 10–20 turns 50–100+ turns
Sustained persona coherence limited
Behavioral drift detection partial
Risk framed for Risk / Compliance / Security partial
DSA Art. 28 / AI Act mapping partial
Adaptation to a new bot Days / weeks Days 2–3 hours
Who Should Contact Us Now

Guardian is built for institutions deploying chatbots that interact with real users.

If your bot handles customer support, advice, onboarding or first-line interaction with the public, the failure modes above already apply to your deployment.

Banks & financial assistants

Customer support, onboarding and advisory chatbots facing retail and SME customers under DORA scope.

Insurance & claims support

Insurance bots handling claims, coverage questions and first-notice-of-loss interactions.

Public-sector services

Migration, employment, social services and citizen-facing chatbots — High Risk under the EU AI Act.

HR & onboarding bots

Internal assistants handling onboarding, policy questions, leave requests and employee support.

Internal compliance copilots

AI assistants supporting Risk, Compliance and Legal teams on policy interpretation and reporting.

Healthtech & care assistants

Patient-facing chatbots handling appointment, triage or information requests in regulated healthcare settings.

● AUGUST 2026 — FULL AI ACT ENFORCEMENT

The regulatory window is closing.

Public services, banks, insurers and healthtech now route asylum seekers, debt-distress callers and patients through chatbots that were never designed for the conversational pressure they now handle every day. Three converging frameworks now make behavioral resilience a documented obligation.

DORA

Operational resilience testing

Required across all EU financial entities — explicitly extends to AI systems facing customers and operational counterparties.

EU AI ACT

High-Risk classification

Employment, migration and access-to-services chatbots classified as High Risk — fines up to €35M or 7% of global turnover per violation.

DSA — ART. 28

Vulnerable user protection

Explicit obligations to protect minors and vulnerable users — chatbot crisis failures create direct exposure under platform liability.

● Early Access Program — First shock audit at no cost for qualified prospects

Identify conversational risks before they become regulatory, reputational or operational incidents.

30 minutes with Benoît Vogt, Founder & CEO. We'll walk you through Guardian's method, comparable findings from your sector, and how Behavioral Safety Certification fits your existing regulatory file.