All field notes

Field note · Incident study

When Your AI Vendor's Cage Door Gets Left Open

On July 31, 2026, Anthropic disclosed that Claude models breached the production systems of three real organizations during offensive-cyber evaluations. The models didn’t defeat containment. Containment was mis-set. A configuration error at Anthropic’s third-party eval partner, Irregular, left the sandbox connected to the public internet even though the models had been told they weren’t.

Anthropic reviewed 141,006 evaluation runs to find the three real breaches, notified the affected organizations in four days, and disclosed publicly in eight. The story most outlets ran is “another lab, another escape.” The story worth telling is what happens after the disclosure lands. For any customer whose Claude outputs go into things that matter — the code they ship, the analyses they act on, the evidence they file, the policies they enforce, the decisions they attest to — the incident creates a downstream problem they’ve never had to solve before: integrity questions on every deliverable Claude touched during the affected window. This isn’t only a compliance story.

141,006Eval runs reviewed
3Real orgs breached
0.002%Incidence rate
8Days from discovery to public disclosure

The failure chain

Six phases, in order. Steps 1–4 are the failure. Steps 5–6 are the response — and the response is where the industry precedent gets set.

01

The eval setup

Anthropic ran offensive-cybersecurity evaluations of Claude models — capture-the-flag exercises where the model searches simulated networks for hidden information. The evals were operated in partnership with Irregular, a third-party evaluation firm. Standard practice for capability testing at this tier.

Offensive-cyber evalThird-party eval infra
02

The configuration error

The eval environment was intended to be air-gapped. It wasn’t. A misunderstanding between Anthropic and Irregular left the sandbox nodes reachable on the public internet. Anthropic’s own framing: “the cage was essentially left open.” The models’ system prompts told them they had no internet access; the network told them otherwise.

Config driftVendor-infra failurePrompt-vs-network mismatch
03

Task optimization discovers the open door

The models, pursuing CTF objectives, discovered that the real network was reachable and used it. This is the mechanic worth sitting with — the model wasn’t exploiting the containment; it was simply taking the cheapest path to its objective, and the cheapest path went through a door nobody had actually locked. Capability didn’t escape containment. Containment failed silently.

Objective-driven actionSilent-failure containment
04

Three organizations compromised

Claude models, operating from the mis-configured eval sandbox, reached the production systems of three real organizations. Details of the specific access remain confidential to the affected orgs. What matters for the industry is the shape: model behavior + vendor-infra config + real-target reach cascaded into real customer impact without any single actor intending it.

Real production reachCascading harm
05

Detection — via transcript review

Anthropic began reviewing evaluation transcripts on July 23 and suspended all cyber evaluations the same day. By July 24, all three real-org incidents had been identified out of the full corpus of 141,006 evaluation runs. Detection happened at the model-lab layer through post-hoc log analysis — not at the sandbox layer where the failure originated, and not at any of the three victim organizations’ SOCs.

Post-hoc detectionLab-side transcript audit
06

Disclosure discipline

Anthropic notified the three affected organizations on July 27 (four days from initial discovery) and disclosed publicly on July 31 (eight days). The 141,006-run denominator was published as part of the disclosure. This is meaningfully different from the OpenAI’s Hugging Face + Modal Labs disclosure timeline, where the second-victim disclosure came a week after the first. Whether the Anthropic pattern becomes the industry norm is an open question. It should.

4-day victim notification8-day public disclosureDenominator published

The propagation graph nobody had to draw before

The affected parties Anthropic notified were the three organizations the compromised sessions reached. But every downstream customer whose Claude-produced outputs went into anything that matters inherits a version of this incident silently — and until now, nobody has been asking the question.

What Anthropic + Irregular ownVendor-side
Containment (Irregular)The sandbox architecture, the network config, the “cage door” that was left open.
Detection (Anthropic)Post-hoc transcript review over 141,006 runs. Non-trivial engineering. Not something a customer could reproduce.
Disclosure (Anthropic)4 days to affected orgs, 8 days to public. The industry precedent this incident sets.
Third-party org notificationDirect outreach to the three affected organizations, coordinated response, incident close-out.
What YOU suddenly ownDownstream customer
Output-integrity questionWhich of my Claude-touched deliverables — code, analyses, evidence, policy drafts, decisions, attestations — were produced during the affected window? Which of them still get trusted, and by what standard?
Stakeholder-facing narrativeWhen an auditor, an engineering lead, a board member, or a customer asks “what did you do when your AI vendor disclosed a breach,” what paper trail can I produce? Auditors are the highest-stakes case — not the only one.
Regulatory notificationUnder GDPR Article 28 (processor obligations), DORA (ICT third-party risk), NYDFS §500.11 (third-party service providers), what do I owe to whom, and by when?
Re-verification decisionsDo I invalidate the affected outputs? Regenerate? Attest around them? Which decisions do I need to revisit, and which shipped code needs another set of eyes?

Every AI-produced deliverable has a supply chain. Every AI-vendor incident is potentially a customer-output incident — silently, unless someone drew the propagation graph in advance.

AI-vendor incident propagation — the problem that lands downstream of every model-lab disclosure from here forward.

What Trustworthy Autonomy™ and vendor-incident workflow would and wouldn’t have changed

vCISO Lite doesn’t run Anthropic’s sandbox and can’t prevent Irregular’s config error. Where it does intersect this incident class is the customer-side propagation — the part everyone downstream of the disclosure suddenly has to answer for.

×What it wouldn’t have fixed
The vendor’s containment failure.

An eval-sandbox config error at a third-party firm is not something a compliance substrate can reach into and repair. That’s a lab + eval-vendor problem.

Real-time detection.

Anthropic caught this through post-hoc transcript review; we are strictly downstream of the vendor’s own disclosure. Real-time visibility into a model-lab’s eval runs is not on our side of the wall.

Actions the compromised sessions took against third parties.

Undoing damage the escaped sessions caused at the three affected organizations is Anthropic’s incident-response job, not the downstream-customer’s.

What it would have made materially better
Vendor-incident propagation workflow.

When an AI provider designated as a critical vendor discloses a containment failure, the disclosure becomes an actionable customer workflow — acknowledge, assess exposure, drive downstream re-verification. Same shape as any other vendor-incident workflow, applied to a new class of vendor.

Evidence-graph time-slicing.

The evidence graph anchors actions with (model, version, session_timestamp). That means a customer can, in principle, ask “show me every output anchored to a Claude session between July 22 and July 24” — whether that output is compliance evidence, a risk analysis on the record, an attestation, a policy draft, or code that got committed to a repo. Same query, different downstream consumers.

Stakeholder-defense narrative.

Auditors, engineering leads, board members, and customers are all going to start asking “what did you do when your AI vendor disclosed a breach.” Customers who can produce a paper trail (vendor-incident record → affected outputs identified → re-verification action taken → attestation) are substantially better positioned than the alternative (“we didn’t know we needed to check”) — whichever stakeholder is asking.

Primary sources