Book Your CGST Assessment

Six-layer governance check. Scored output with remediation roadmap. 2 hours. Run it before someone else does.

Book the Assessment →

In April 2026, Anthropic disclosed a finding that should be on every enterprise AI team’s radar: Mythos, an AI system capable of autonomously identifying zero-day vulnerabilities and writing production-quality exploits, at low cost. The specific run that found one OpenBSD vulnerability cost under $50; Anthropic put the total cost of that search, about a thousand runs that also produced several dozen other findings, at under $20,000.

The economics of autonomous AI capability have changed. What took a skilled red-team hours now takes an AI agent minutes. And the governance question that follows is not theoretical: if an adversary can search for vulnerabilities at that cost, how confident are you in the governance architecture of your own autonomous agents?

The CGST — Constitutional Governance Stress Test — answers that question with data.

The Governance Gap That Audits Keep Missing

Most enterprise AI governance programs are built around access and compliance. IAM policies, audit logs, approved vendor lists, documented use cases. These are necessary. They are also structurally incomplete.

Grant Thornton’s 2026 AI Impact Survey of nearly 1,000 US business leaders found that 78% lack full confidence that their organization could pass an independent AI governance audit within 90 days. Many governance programs evaluate access, not action. They confirm that an agent has the right credentials, not that the agent makes good decisions once it has them. An authorized agent with no behavioral governance can still fabricate data, overspend budgets, operate outside its intended scope, and fail silently for weeks.

That is the gap the CGST is built to expose.

78%
of senior leaders lack full confidence they could pass an independent AI governance audit within 90 days
Grant Thornton 2026 AI Impact Survey
<$50
Cost of the single Mythos run that found one zero-day; the full search cost under $20,000
Anthropic disclosure, April 2026

What the CGST Actually Tests

The CGST runs your agent system through six evaluation layers drawn from the same constitutional governance architecture we have operated in production for 107+ days. Each layer has a specific failure mode it is designed to catch.

EG
Epistemic Gate — Can your agents be certain enough before acting, or do they fabricate confidence? Tests for data hallucination, scope creep, and self-reported authorization.
RG
Risk Gate — Do your agents evaluate reputational and trust risk before taking action, or do they execute first and flag later? Tests for brand-damaging outputs sent without review.
GG
Governance Gate — Are your agents gaming their own metrics? Tests for Goodhart’s Law failures: optimizing measured KPIs in ways that harm unmeasured outcomes.
EPG
Economic Performance Gate — Do your agents have spending constraints that cannot be overridden? Tests for unbounded cost creation and budget bypass paths.
AAG
Autonomy Assurance Gate — Are your agents operating within defined authority levels, or autonomously escalating into decisions that require human review? Tests for scope expansion and authority creep.
CGG
Constitutional Governance Gate — Can your agents detect and prevent unauthorized self-modification? Tests for capability expansion that bypasses governance controls.

Each gate produces a PASS, HOLD, or FAIL finding with a specific remediation path. The final output is a governance score on a 100-point scale — with the ungoverned baseline for reference. Our own constitutional-agent library scored 63/100 governed vs. 6/100 ungoverned. The delta tells you exactly what governance is buying you.

Why This Is a Revenue Question, Not Just a Risk Question

Governance also matters for regulatory readiness: most obligations of the EU AI Act apply from August 2, 2026.

If you are selling to enterprises, your buyers’ procurement teams are already asking governance questions. Whether your AI systems operate under hard constraints, whether you have an audit trail for agent decisions, whether your system has been stress-tested against adversarial inputs — these are now procurement gates, not nice-to-haves.

The CGST produces a scored report you can show to enterprise buyers. Not a narrative about your commitment to responsible AI. A number, with methodology, with gap analysis, with remediation evidence.

The Cheap Zero-Day Reframe

A single successful Mythos run cost under $50, and the whole search that produced it cost under $20,000. Either number is small next to the cost of a breach. When finding weaknesses gets that cheap, the question is no longer whether a sophisticated adversary could find your governance gaps, but whether you find them first.

What the Assessment Looks Like

The CGST is a structured 2-hour engagement. We evaluate your agent architecture against the six-gate framework, review your existing governance controls, run adversarial probes against the specific failure modes each gate is designed to catch, and produce a scored report with prioritized remediation.

You get a governance score. You get a layer-by-layer breakdown of where your system passes, where it holds, and where it fails. You get a remediation roadmap ordered by risk severity. And you get the methodology — the same one formalized in our published preprints — so your team understands why each finding matters.

Two tiers are available depending on your organization’s scale and urgency:

Tier 1 Assessment
$299
Six-gate evaluation against your documented architecture. Scored report, gap analysis, and prioritized remediation roadmap. Best for teams that want a fast governance baseline before a procurement conversation or regulatory review.
Book Tier 1 →
Tier 2 Assessment
$2,000
Full adversarial stress test with live agent interaction, OWASP ASI Top 10 coverage, custom threat modeling for your deployment context, and executive-ready remediation brief. Best for organizations with production agent deployments and active enterprise procurement or EU AI Act compliance timelines.
Book Tier 2 →

The Window for Self-Assessment Is Closing

Most obligations of the EU AI Act apply from August 2, 2026. NIST has active listening sessions on AI agent standards. Singapore published the first governmental framework for agentic AI governance. These frameworks describe behavioral requirements — transparency, human oversight, risk management, robustness under adversarial conditions. Compliance programs built on access governance alone will not satisfy them.

More immediately: if Mythos-class tools can search for weaknesses this cheaply, the adversarial pressure on production AI systems will increase, not decrease, over the next 12 months. Organizations that have stress-tested their governance before that pressure arrives will have a structural advantage over those that learn about their gaps from an incident report.

The organizations that run their own governance stress tests before deployment will spend weeks on remediation. The organizations that don’t will spend months on incident response — at costs that dwarf what a pre-deployment assessment would have found.

The CGST is not a compliance checkbox. It is a production readiness test for systems that are already making consequential decisions autonomously. Run it now, while the findings are information rather than liability.

Book Your Governance Stress Test

Tier 1 ($299) or Tier 2 ($2K). Scored output. Six layers. Remediation roadmap. Run it before a bad actor does.

Book the CGST →

Read the Underlying Research

The six-gate architecture is formalized in two published (not peer-reviewed) preprints: Constitutional Self-Governance (12 mechanisms, NIST/EU AI Act mapping) and the Agent Security Harness (342 tests, OWASP ASI Top 10 coverage).

Constitutional Self-Governance (Zenodo)    Agent Security Harness (Zenodo)

Frequently Asked Questions

How is the CGST different from a security penetration test?

A pen test evaluates access control and vulnerability exposure. The CGST evaluates governance architecture — whether your agents operate under enforceable behavioral constraints, whether those constraints hold under adversarial pressure, and where decision-making authority is improperly bounded. The two are complementary, not equivalent. Your pen tester confirms the front door is locked. The CGST confirms the agent inside the building has a constitution.

What does the scored report include?

A governance score on a 100-point scale, a gate-by-gate breakdown of PASS/HOLD/FAIL findings, specific failure modes identified with evidence, and a prioritized remediation roadmap. Tier 2 includes an executive brief formatted for board presentation and enterprise procurement review.

Does this apply if my agents use off-the-shelf platforms like Microsoft Copilot or Salesforce Agentforce?

Yes. Platform-level governance controls what agents can access. The CGST evaluates what your agents do with that access — the decision layer that platform IAM does not govern. The CGST is designed to run regardless of underlying platform.

Related Articles

Corrections, 28 September 2026: We corrected the Mythos cost figure: Anthropic reported that the one run that found a particular OpenBSD vulnerability cost under $50, and that the full search cost under $20,000; it did not report $50 per zero-day. We corrected the Grant Thornton figure, which measures leaders' confidence in passing an AI governance audit, not the content of audit frameworks. We removed a claimed PwC “7.2x return on governance investment” and the related 40% and $2.1M figures, which we could not find in PwC's research; PwC's 7.2x figure compares AI-driven gains between companies. We corrected “peer-reviewed preprints”: the preprints are published but not peer reviewed.

Get notified about new EAA series parts

One email when a new position-paper part or field note publishes. No newsletter, no marketing list.