Mythos: free retest with each new frontier model | Swarm
Frontier-model commitment|Sign in
27
YEARS HIDDEN  /  OPENBSD'S SACK HANDLERMythos found it in weeks

For 27 years,OpenBSD's SACK handlershipped a bug nobody found.

On April 7, 2026, Anthropic announced Claude Mythos Preview — a model described as "strikingly capable at computer security tasks." Within weeks of release, it surfaced critical bugs across FreeBSD, OpenBSD, Firefox, and the Botan cryptography library. Three decades of adversarial review never found them.

Swarm rides that wave. Every paying customer gets a free re-test with each new frontier model we onboard.

UK AI Security InstituteEvaluation · Apr 13, 2026
73%
Expert-level CTF
3/10
32-step attack chain
181×2
Firefox exploits vs Opus 4.6

The dossier

Three claims. One citation each.

Every number on this page traces to a public source. Read the citation. Decide for yourself.

  1. 01
    UK AI Security Institute · Apr 13, 2026

    73% expert-level CTF success rate

    On capture-the-flag tasks no model could complete before April 2025, Mythos Preview solved 73% of expert-level challenges. AISI's numbers, not ours.

    CITATIONaisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
  2. 02
    AISI "The Last Ones" evaluation

    First model to complete the 32-step enterprise attack chain

    Reconnaissance, credential theft, web exploitation, reverse engineering, CI/CD compromise, SQL injection — start to finish, three runs in ten.

    CITATIONaisi.gov.uk/blog/our-evaluation-of-claude-mythos-previews-cyber-capabilities
  3. 03
    Anthropic Mythos Preview · Apr 7, 2026

    181 working Firefox exploits versus 2 from Claude Opus 4.6

    Same vulnerabilities in Mozilla Firefox 147's JavaScript engine. Mythos produced 181 working exploits. The previous-generation model produced 2.

    CITATIONred.anthropic.com/2026/mythos-preview

Disclosed findings

Bugs that hid for decades.

In weeks of red-team use, Mythos Preview surfaced critical-severity flaws across some of the most adversarially-audited code in open source.

  1. 27yrs
    OpenBSD
    TCP SACK handler · DoS
    A denial-of-service flaw that sat in one of the most adversarially-audited codebases in open source for 27 years. Crashes any OpenBSD host that responds over TCP.
  2. 0day
    FreeBSD
    CVE-2026-4747 · NFS RCE
    Unauthenticated remote code execution against the NFS daemon. Complete server takeover from any attacker on the internet.
    CVE record
  3. Crit
    Botan
    Certificate auth bypass
    A flaw in one of the most widely-used C++ cryptography libraries that lets certificate-based authentication be bypassed entirely.
The standing offer

Every paying customer gets a free re-test with each new frontier model we onboard.

Why we can commit

The orchestrator is model-agnostic. Upgrading from Opus 4.7 to the next frontier model is a configuration change, not a rebuild. Every specialist inherits the upgrade.

What it covers

Mythos Preview, any successor Anthropic releases more broadly, and any comparably-capable model from another lab. Standing commitment on every active contract.

Glasswing contextAWS · Apple · Cisco · CrowdStrike · Google · JPMorganChase · Linux Foundation · Microsoft · Palo Alto Networks
References

Sources

  1. 01AISI · Our evaluation of Claude Mythos Preview cyber capabilities (Apr 13, 2026)research
  2. 02Anthropic · Claude Mythos Preview (Apr 7, 2026)research
  3. 03CVE-2026-4747 · FreeBSD NFS RCE (NVD record)cve
Read the receipts.
ENTER YOUR DOMAIN. SWARM MAPS YOUR ATTACK SURFACE IN JUST A FEW MINUTES.No card. Free preview.