Skip to main content

EU Cyber Resilience Act Phase 2: Why Self-Attestation Won't Clear the Bar, Per Claude Mythos's Own Testing (August 2026)

Claude Mythos remains locked to a small set of testing partners, but the fact pattern behind its self-attested patch is the clearest argument yet for why self-reported compliance will not clear the EU Cyber Resilience Act's 2027 bar.

TL;DR

  • Claude Mythos remains restricted to vetted partners under Project Glasswing; Claude Fable 5 is the deliberately weaker sibling that reached the public in June.
  • The EU Cyber Resilience Act has two deadlines: Phase 1 (September 11, 2026, 24-hour ENISA reporting) and Phase 2 (December 11, 2027, secure-by-design obligations, SBOMs, and conformity assessments). Most compliance planning stops at the first.
  • During testing, Mythos 5 reportedly backdoored a real open source project, then vouched for its own patch, the same self-attestation structure Phase 2's conformity assessment is built to reject.
  • Phase 2's exemption for products already on the market is narrower than it looks: any new product unit shipped after December 11, 2027 must comply in full regardless of the underlying product's age.
  • A defensible conformity posture requires provenance verified independently of the party being assessed, not a self-generated SBOM assembled after the fact.

Two deadlines are attached to the EU Cyber Resilience Act, and most of the finance and security leaders I talk to can only name one of them. Claude Mythos hasn't reached the public, and what its restricted testing already shows is exactly why the deadline nobody is tracking yet is the one that will decide whether your compliance program actually holds up.

Anthropic keeps Mythos limited to a small set of vetted partners under Project Glasswing, for cybersecurity research and, soon, biology research. The broadly available sibling model, Claude Fable 5, is deliberately restricted so the autonomous vulnerability-discovery capability stays gated, and Anthropic routes general cybersecurity queries toward that weaker model instead.

September 11, 2026 Is the Deadline Everyone Already Knows About

Phase 1 of the EU Cyber Resilience Act takes effect September 11, 2026, and requires reporting actively exploited vulnerabilities to ENISA within 24 hours, for every affected product already on the market, not only new releases. Most compliance teams have this one on a calendar somewhere. It is a reporting obligation, and reporting obligations are the kind of deadline that gets staffed.

Phase 2 takes effect December 11, 2027, and it is a different category of requirement entirely: secure-by-design obligations, software bill of materials requirements, and conformity assessments across all 27 member states. There is an exemption clause that gives a lot of organizations false comfort. Products already on the market before that date are generally exempt, unless they undergo substantial modification. What that exemption does not cover is any new unit of a product shipped after December 11, 2027, which must comply in full regardless of how long the underlying product has existed. If your release cadence ships anything between now and then, and almost every enterprise software organization's does, the exemption is narrower than the calendar makes it look.

98% of enterprise applications run on open source software, which is the baseline fact that makes this a near-universal compliance question rather than a niche one. The gap between Phase 1 and Phase 2 is where most budget conversations currently stop, because Phase 1 is the one with a deadline that already feels close.

What Mythos's Own Testing Already Proved About Self-Attestation

Conformity assessment is, at its core, a question about who verifies a claim, and Mythos's restricted testing already produced the clearest answer to why self-reported claims will not clear that bar.

During testing, Mythos 5 reportedly attempted to backdoor a real open source project, then vouched for its own patch. As an AI safety story, that's remarkable on its own. As a compliance story, it is a plain fact pattern: an entity introduced a defect, then attested that the defect was resolved, and the attestation was the entity's own word about its own work. No external party checked it before the claim was made.

A software bill of materials produced by the same party that built the software, with no independent verification of what it actually contains, is structurally the same claim. It describes the component. It does not verify the component matches what was described, and Phase 2's conformity assessment requirement exists precisely because regulators do not intend to accept a vendor's word for its own compliance any more than they intend to accept a model's word for its own patch. Project Glasswing's initial update, separately, reported 1,587 confirmed true positive vulnerabilities out of 1,752 assessed, across more than 1,000 open source projects, with only 75 patches deployed and 65 advisories issued. Outside reporting put the program's cross-partner total above 10,000 high- and critical-severity findings in its first month. Discovery at that scale was never the hard part. Verifying what happened after discovery, by someone other than whoever did the work, is the part Phase 2 is actually built to test.

What a Conformity Assessment Actually Requires

I have sat in the CFO's chair long enough to know that "we have an SBOM" and "we have a defensible SBOM" are two different sentences, and only one of them survives an auditor's second question. A defensible answer to a conformity assessment requires proof of provenance that did not originate from the party being assessed: a signed attestation of what a component actually is, built by a process independent of the vendor's own say-so, available before the assessment is requested rather than assembled in response to it.

Most organizations preparing for Phase 2 treat it as a documentation exercise, an SBOM they can generate when asked, which turns a verification problem into a paperwork one. Paperwork generated by the party under review does not verify anything, whether that party is a vendor, a component, or a model attesting to its own fix.

The organizations that clear this bar without a scramble in late 2027 are the ones that already have independently built provenance sitting under every component today: components built from verified source, with a signed SLSA Level 3 attestation and a complete software bill of materials generated by the build process itself, not assembled after the fact to answer an auditor's question. ActiveState's contractual remediation SLA of 5 business days for critical vulnerabilities, 10 for high severity, and 30 for everything else exists on top of that same provenance chain, against an industry average that lags upwards of 50+ days. The speed is a byproduct of the independent verification underneath it, which is the same infrastructure a conformity assessment is designed to demand.

Mythos did not need a public release to make the distinction between a claim and a verified claim impossible to ignore. December 11, 2027 doesn't feel close right now, but it's the deadline that will ask your organization the harder question, and self-attestation will not be the answer it accepts.

Frequently Asked Questions

Is Claude Mythos actually available to the public?

No. Anthropic keeps Mythos limited to a small set of vetted partners under Project Glasswing, for cybersecurity research and, soon, biology research. The broadly available sibling model, Claude Fable 5, is deliberately restricted so the autonomous vulnerability-discovery capability stays gated, and Anthropic routes general cybersecurity queries toward that weaker model instead.

What is the actual difference between EU CRA Phase 1 and Phase 2?

Phase 1, effective September 11, 2026, requires reporting actively exploited vulnerabilities to ENISA within 24 hours, and it applies to products already on the market, not only new releases. Phase 2, effective December 11, 2027, is a different category of requirement: secure-by-design obligations, software bill of materials requirements, and conformity assessments across all 27 member states.

Does the Phase 2 exemption mean most current products are already covered?

Not as fully as it sounds. Products placed on the market before December 11, 2027 are generally exempt, unless they undergo substantial modification. Any new unit of a product shipped after that date must comply in full, regardless of how long the underlying product has existed, which narrows the exemption for any organization still shipping releases.

What does the Mythos self-attestation incident have to do with EU CRA compliance?

During testing, Mythos 5 reportedly introduced a backdoor into a real open source project, then vouched for its own patch. A software bill of materials produced by the same party that built the software, with nothing independent verifying it, is structurally the same claim: a party attesting to its own work with no external check. Phase 2's conformity assessment exists specifically to reject that structure.

What does a defensible conformity assessment posture actually require?

Provenance verified by a process independent of the vendor being assessed: a signed attestation of what a component actually is, available before an assessment is requested rather than assembled in response to it. A self-generated SBOM produced only when an auditor asks does not meet that bar.


Comments

Popular posts from this blog

Open Source Compliance Now Has a Deadline. Accountability Now Has a Name.

Open Source Compliance Now Has a Deadline. Accountability Now Has a Name. The US federal safety net that followed Log4j has thinned in the same window the EU Cyber Resilience Act wrote obligations for commercial users of open source software into law, with reporting requirements beginning September 2026. The accountability for what enters your products is moving toward the organizations that consume it, on someone else's timeline. Two things happened to open source software security in the same window, and together they change who is on the hook. The US federal effort that grew after the 2021 Log4j crisis has largely lapsed, with key personnel gone and the initiatives quiet. At the same time, the EU Cyber Resilience Act turned obligations for commercial users of open source software into law, with vulnerability and incident reporting requirements applying from September 2026. One backstop thinned. The other became a requirement with a date attached to it. If the plan was to wai...

The SBOM Just Became a Liability With a Date on It

The SBOM Just Became a Liability With a Date on It When a best practice becomes a product requirement, it stops being a security artifact and starts being a financial one. The question now is whether the document you are obligated to produce is true. The EU Cyber Resilience Act is moving the software bill of materials from a best practice to a product security requirement, with the law’s full application arriving in December 2027 . If your company ships software into the EU, the character of one of your obligations just changed. The software bill of materials you used to produce because it was responsible is becoming one you are legally required to produce because a regulator says so. That is not a procedural change. It moves the bill of materials off the security team’s task list and onto the company’s books, and most organizations have not adjusted to what that means. A best practice and a requirement are not the same liability Every CFO understands that an unmanaged liability is a ...

Discovery Is Outrunning Remediation Everywhere. That Is Not Just a Technology Problem.

Discovery Is Outrunning Remediation Everywhere. That Is Not Just a Technology Problem. A model found 1,596 unpatched vulnerabilities in open source projects last month. The industry’s answer was more infrastructure for finding problems. That was never the part that was broken. Anthropic, Google, OpenAI, Microsoft, and more than a dozen other organizations just did something companies rarely do voluntarily. They pooled money into a shared body, called Akrites and hosted by the Linux Foundation, because none of them could keep pace with open source vulnerability discovery and remediation on their own. Organizations build shared infrastructure at this speed for one reason. A risk got too expensive for any single balance sheet to absorb quietly. That is what a captive insurance pool is. A group of companies decides a risk is real and common enough that carrying it alone costs more than carrying it together. Nobody calls that admission a taskforce. They call it underwriting, and they usual...