The External Referent Arrived: What the Curl Test Tells Us About Vendor Verification
Previous: Episode 16: When the Tool Found the Bug | Next: Episode 18: First Implementation Attempt | GitHub: Te Pono V.E.R.A.
Published: 23 May 2026 | Series: Real Intelligent AI | GitHub: https://github.com/andreas-linux/vera/
Five "confirmed vulnerabilities." One confirmed. Three false positives. One non-security bug. An 80% rejection rate. Reviewed by the people who know the codebase better than any model ever will.
The Question the Coding Arc Has Been Building Towards
Episode 15 closed with a design brief. The argument there was that a verification harness cannot verify itself: KAIROS, the canonical worked example for the Te Pono V.E.R.A. Logic Engine, requires an external referent with its own audit history, sitting outside the system being evaluated. The closing question was precise. What external referent would KAIROS's consolidation process need, and where would it live?
Episode 16 sharpened the structural urgency. CopyFail (CVE-2026-31431), surfaced by Theori's Xint Code AI tool in approximately one hour after nine years undetected, established that AI-assisted discovery can collapse the cost of finding a deep logic flaw in foundational system software. What did not collapse was the verification cost. The gap between discovering something and establishing that the discovery is correct, complete, and not a false positive remained structurally unchanged.
Six episodes into the Coding Arc, the empirical test arrived.
On 11 May 2026, Daniel Stenberg, curl lead developer, published findings from a third-party assessment of Anthropic's Claude Mythos model against curl's codebase. This is the external-referent event the arc has been building towards.
What Happened on 11 May 2026
Claude Mythos is Anthropic's restricted-access offensive cyber AI model, distributed to a limited number of major organisations under Project Glasswing, administered through the Linux Foundation. At launch, Anthropic positioned Mythos as having identified thousands of zero-day vulnerabilities prior to release. The model represents Anthropic's direct entry into frontier AI-assisted vulnerability discovery.
Stenberg did not have direct access to Mythos. A third party with Project Glasswing access ran the model against curl's git repository and provided the resulting report to Stenberg. The scan covered approximately 176,000 lines of C code at a recent master-branch commit.
The initial report listed five confirmed security vulnerabilities. After review by Stenberg and the curl security team, the list was revised:
One confirmed, low-severity vulnerability. Scheduled for the curl 8.21.0 release in late June 2026.
Three false positives, each linked to curl limitations already documented in the API documentation.
One ordinary bug with no security implication.
Stenberg's conclusion was direct. The hype surrounding Mythos was, in his framing, "primarily marketing." He noted explicitly that he saw no evidence the model finds security issues to any meaningfully higher degree than existing tools.
The table below consolidates the key facts from the assessment.
| Item | Detail |
|---|---|
| Date of publication | 11 May 2026 |
| Model tested | Anthropic Claude Mythos (restricted access, Project Glasswing) |
| Codebase | curl -- approximately 176,000 lines of C |
| Audit history of codebase | OSS-Fuzz, Coverity, CodeQL, multiple paid audits (per Mythos report) |
| Access method | Third party with Project Glasswing access ran scan; provided report to Stenberg |
| Initial findings reported | 5 "confirmed security vulnerabilities" |
| Findings after external review | 1 confirmed (low severity); 3 false positives (documented API limitations); 1 non-security bug |
| Confirmed CVE status | Low-severity; scheduled for curl 8.21.0 release, late June 2026 |
| Stenberg conclusion | Hype around Mythos was "primarily marketing"; no evidence of meaningfully higher detection than existing tools |
| External referent | Stenberg and the curl security team, with full codebase audit history |
Table 1: Stenberg / Mythos / curl assessment: key facts. Sources: Stenberg blog (11 May 2026); The Register; SecurityWeek; SC Media.
The External Referent Did Its Work
The structural argument in Episode 15 was not that external referents always produce surprising findings. It was that without an external referent, there is no mechanism by which a vendor capability claim can be distinguished from a verified capability finding. The Stenberg assessment is the demonstration.
The curl security team is not a neutral observer who happened to show up. The codebase itself carries its verification history. Curl has been processed through OSS-Fuzz continuously. It has been reviewed by Coverity and CodeQL. Multiple paid audits have examined it over its lifetime. The Mythos assessment report acknowledged this directly: curl is one of the most fuzzed and audited C codebases in existence.
This is the point. The external referent's credibility derives from its own prior work, not from the model's output. When Stenberg and the curl security team reviewed five reported findings and confirmed one, they were not guessing. They were applying accumulated expert knowledge to a set of machine-generated claims. That is what external-referent verification means in practice.
The design brief from Episode 15 asked where the external referent would live and what it would need. The curl case answers it concretely. The referent lives in the codebase's audit history, in the expertise of the team that has maintained it, and in the institutional memory of a project that has faced serious security review for years. You cannot construct that referent the week before you need it. It accumulates, or it does not exist.
The Calibration Note: Mozilla and the Hardened Codebase Problem
The curl result requires careful reading. It is not evidence that Mythos cannot find vulnerabilities. Mozilla reportedly credited Mythos with identifying more than 270 bugs across the Firefox codebase. Mozilla's security team indicated that every issue the model identified could have been found by skilled human researchers; the value was in the speed of discovery, not in the uniqueness of the findings.
Curl is a different class of codebase. It has been under continuous aggressive fuzzing and formal static analysis for years. The attack surface that remains after OSS-Fuzz, Coverity, CodeQL, and multiple paid audits is not the same as a large, complex browser engine with a broader scope. The Stenberg finding is calibration data: against a maximally hardened target with a long verification history, Mythos's vendor-stated capability and its independently reviewed capability are not the same.
Erik Cabetas of Include Security has reportedly observed separately that multiple organisations granted Mythos access have seen results consistent with the curl finding. The curl result is not an outlier. It is a data point in a pattern.
OpenAI launched Daybreak in the same week, powered by GPT-5.5-Cyber and Codex Security, with access similarly restricted to trusted defenders for critical infrastructure. Two frontier labs now compete on dual-use cyber AI with restricted-access governance structures, and neither has yet been independently validated at scale using the methodology Stenberg applied. The curl test is not just a finding about Mythos. It is the template for what independent assessment looks like.
The Verification Gap Framework Update
The Verification Gap framework, introduced in Episode 7 and grounded empirically in Episode 16, now has its strongest external anchor. The table below maps the theoretical position against the empirical finding.
| Dimension | Episode 7 (theory) | Episode 17 (empirical) |
|---|---|---|
| Vendor claim | Assumed at face value in most enterprise deployments | Anthropic: "thousands of zero-days" against Mythos pre-launch |
| Verified capability | Hypothetically divergent from vendor claim under external review | Curl result: 1 of 5 confirmed; calibrated against hardened codebase |
| External referent | Required by architecture; not yet demonstrated in practice | Stenberg / curl security team; own audit history established independently |
| Gap status | Theoretical; NTP-derived architectural principle | Empirically observed; strongest external anchor to date |
| Governance implication | External referent necessary, not optional | External referent methodology demonstrated; replicable for future assessments |
Table 2: Verification Gap: theoretical position (Episode 7) vs empirical anchor (Episode 17). Sources: Te Pono V.E.R.A. architecture; Stenberg blog; SecurityWeek.
What the Curl Test Validates About Te Pono V.E.R.A.
The Te Pono V.E.R.A. Logic Engine's Verification Gap framework now has its strongest external empirical anchor. For six Coding Arc episodes, the argument has been structural and theoretical, grounded in Non-Traditional Predication Theory and in architectural analysis of existing systems. The Stenberg assessment is the first widely-publicised independent test of a frontier cyber AI model's stated capability against a production codebase with a deep audit history. It directly confirms the architectural premise.
The Heaven Vector in this event is Stenberg's methodology: a transparent, peer-reviewed, publicly reported independent assessment, conducted by a named team with established domain expertise, against a codebase with a documented verification history. This is what verification governance looks like when it operates correctly. The Skynet Vector is the prior state: a vendor capability claim, accepted without independent assessment, distributed under restricted access to organisations that lacked the referent infrastructure to evaluate it independently.
The architectural principle the Te Pono V.E.R.A. Logic Engine is built around is the separation of predication from existence: the formal NTP distinction between describing a property and asserting that something actually obtains. "Mythos found 5 confirmed vulnerabilities" is an existence assertion. It requires independent verification before it can be treated as a fact. Stenberg's review is that verification. The revised number, one, is what the existence assertion resolves to under external-referent review.
That is not a failure of AI. It is a demonstration of why verification architecture is not optional.
What Te Pono V.E.R.A. Is Not Claiming
This section is deliberate. The Te Pono V.E.R.A. Logic Engine has a hard exclusion list, and it applies here.
Te Pono V.E.R.A. would not have found the confirmed curl vulnerability. The V.E.R.A. Logic Engine is a claim verification architecture grounded in NTP. It classifies claims, routes them to verification, and assesses existence through formal logical structure. It is not a vulnerability scanner, not a code analysis tool, and not a replacement for the kind of deep domain expertise that Stenberg and the curl security team bring to bear on their codebase.
The V.E.R.A. Logic Engine cannot verify Mythos's reasoning. It cannot evaluate the internal inference process of another model. The claim that Mythos identified a vulnerability is an existence assertion that can be checked against an external referent; the reasoning path Mythos used to arrive at that finding is not accessible to external verification of the kind the V.E.R.A. Logic Engine performs.
What Episode 17 is about is the architectural requirement: that vendor capability claims are existence assertions, not n-type logical truths; that they require external-referent verification; and that the referent must carry its own audit history, independent of the vendor being assessed. The Stenberg / curl / Mythos event demonstrates that requirement in operation. It does not establish Te Pono V.E.R.A. as the mechanism for performing it.
Te Pono V.E.R.A.'s contribution is the formal logical basis for knowing why that requirement exists, and the architecture for applying it to natural language claims in enterprise and governance contexts.
The Coding Arc Closes
Episodes 12 through 17 have traced a single argumentative arc. The table below summarises what each episode established.
| Ep | Title | Core argument |
|---|---|---|
| 12 | The Hardest Problem | Claim extraction as the V.E.R.A. Logic Engine's hardest engineering challenge; Focus/Coverage tradeoff; NTP as check-worthiness ontology |
| 13 | The Mechanism Behind the Lie | Training incentive mechanisms produce confident wrong answers as an objective function outcome |
| 14 | Language Is Not Capability | Governance documents cannot audit themselves; LLMs treat governance vocabulary as governance substance |
| 15 | The Harness Is the Moat | A verification harness cannot verify itself; an external referent must sit outside the system being verified |
| 16 | When the Tool Found the Bug | CopyFail (CVE-2026-31431): discovery cost collapsed; verification cost did not. The Verification Gap is now empirically grounded at the deep-logic layer of foundational software |
| 17 | The External Referent Arrived | Stenberg / curl / Mythos test: external referent demonstrates in practice that vendor capability claims require independent assessment with its own audit history |
Table 3: Coding Arc (Episodes 12-17): argument progression. Te Pono V.E.R.A. Article Registry.
The arc that began with claim extraction as an engineering problem closes with an empirical demonstration of why the engineering problem matters. Extracting claims correctly, verifying their existence against an external referent, and maintaining formal separation between what is described and what is asserted: these are not theoretical constructs. The curl test shows them in operation at the frontier of AI capability assessment.
The next arc continues from here. The questions are practical: how does the Te Pono V.E.R.A. Logic Engine's claim extraction module perform against real governance language? Where does the Ffact score stabilise under test? What does the KAIROS verification pipeline look like when it runs end to end? The architectural argument is made. The implementation work continues.
The Structural Call
The Stenberg / curl / Mythos assessment is the methodology. One person, one team, one codebase with an honest audit history, and the willingness to say publicly what the numbers show. Four of five findings rejected. One confirmed, scheduled.
That is what external-referent verification produces. Not a perfect finding, not a definitive refutation, but a calibrated, reviewable, reproducible result that a board, a procurement officer, or a security team can actually use. The alternative is a vendor's self-description, accepted on authority, with no mechanism to distinguish it from a calibrated finding.
Boards forming a view on frontier AI capabilities have guidance available. NCSC NZ guidance on agentic AI and frontier AI products remains the most relevant NZ-side reference. The Five Country Council joint guidance, co-authored by NCSC NZ and published 1 May 2026, identifies 23 agentic AI risks and is the operative cross-government reference for this assessment category. The Stenberg methodology is what applying that guidance looks like in a specific, technical, empirically grounded context.
Build the referent before you need it. Build it from accumulated audit history, domain expertise, and institutional continuity. Then verify against it.
That is the architecture. That is the moat. That is what the arc has been building towards.
Join the Movement
Te Pono V.E.R.A. is an open-source project under GPL-3.0. The Krampitz Load Analyzer and Formula Parser are available on GitHub. The Coding Arc articles trace the engineering and architectural decisions in real time.
GitHub: https://github.com/andreas-linux/vera/
If you work in AI governance, enterprise architecture, security verification, or formal logic, and you see what the Verification Gap means for your organisation: reach out. The verification problem is structural. The solution requires people who understand why that matters.
Ita est momentum veritatis.
Your organisation is acting on vendor AI capability claims right now. What would it take to build the external referent that could actually verify them?
The views expressed in this article are entirely my own, informed by more than 30 years of professional experience in architecture, security, and technology leadership in New Zealand. They do not represent the views of my employer, any government agency, or the New Zealand government. My commentary on legislation and policy is analytical, drawing on publicly available sources and my professional expertise in architecture, security, and AI governance. I follow the Public Service Commissioner's Code of Conduct for the Public Sector and social media guidance.
Andreas Hamberger, M.A. (Humboldt University Berlin), AMInstD, IAPP CIPP/E, TOGAF Practitioner, is a New Zealand-based enterprise architect and security practitioner with more than 30 years of experience in architecture, security, and technology leadership. He is the founder of Te Pono Limited and the creator of the Te Pono V.E.R.A. Logic Engine. V.E.R.A. (Verified Existence and Reasoning Architecture) is an open-source logic engine available on GitHub.
I use AI tools, including Sudowrite, Claude, Perplexity AI, DeepSeek AI, ChatGPT, Grok, Copilot, Openart and Gemini, as deliberate production tools, not ghostwriters. This is consistent with my position: AI amplifies human judgement; it does not replace it. The frameworks, arguments, and editorial decisions in this series are original work. AI accelerated the process. The thinking is mine.
[1] Stenberg, D. "AI and security, yes really." Personal blog. 11 May 2026. https://daniel.haxx.se/blog/
[2] The Register. "Curl lead developer reviews Claude Mythos security findings." May 2026.
[3] SecurityWeek. Coverage of Anthropic Claude Mythos assessment against curl codebase. May 2026.
[4] SC Media. Coverage of Anthropic Claude Mythos / curl assessment results. May 2026.
[5] Mozilla Security Blog. Reported findings from Claude Mythos testing against Firefox codebase. 2026. [Likely; secondary citation.]
[6] Include Security / Cabetas, E. Reported results from Mythos testing across multiple organisations. 2026. [Likely; secondary citation.]
[7] OpenAI. Daybreak launch announcement. May 2026.
[8] NCSC New Zealand. Guidance on agentic AI and frontier AI products. 2026. https://www.ncsc.govt.nz/
[9] Five Country Council. "Joint guidance: agentic AI risks." Co-authored with NCSC NZ. 1 May 2026.
[10] Hamberger, A. "The Verification Gap." V.E.R.A. Saturday, Episode 7. 14 March 2026.
[11] Hamberger, A. "The Harness Is the Moat." V.E.R.A. Saturday, Episode 15. 9 May 2026.
[12] Te Pono Limited. V.E.R.A. Coding Arc Documentation and Article Registry. 2026. https://github.com/andreas-linux/vera/