Enterprise Cybersecurity IntelligenceMonday

An enterprise cybersecurity intelligence company.For security and technology leaders.

Security.io Intelligence

What changed, why it matters,
and how it evolved.

Supply Chain · Executive briefing

Cursor traces poisoned extension to a Fireworks key and secret-bearing agent traces

Cursor's Friday incident update, amplified over the weekend, traced a poisoned Claude Code extension to a Fireworks-exposed key, secret-bearing agent traces and secondary credential compromise.

AI SecuritySupply ChainEndpoint Security
Why it is in today’s brief

The malicious extension appeared on October 29, but Cursor's November 13 update newly linked the event to a Fireworks-exposed key, secret-bearing agent traces and secondary SaaS access. That changed the enterprise decision from checking two package versions to reviewing AI-vendor data flows and credential boundaries. It earns inclusion over narrower malware items because the trust failure crossed provider, stored-data, marketplace and developer-endpoint layers.

Read first

Identify every installation of Anthropic.claude-code versions 2.1.312 and 2.1.313 from Cursor's marketplace, hunt the published SSH, file, process and network indicators, and rotate credentials reachable from positive hosts.

Act now

Query extension telemetry for Anthropic.claude-code versions 2.1.312 and 2.1.313.

Accountable owner

CISO with developer security, endpoint security, identity and AI-platform governance owners

Decision horizon

Immediate endpoint hunting and credential containment today; complete AI-vendor data-flow and marketplace-control review within seven days.

AssessmentHigh confidence
Emerging riskAdditional tampered extensions, attacker activity beyond Cursor's stated containment point, expanded trace exposure or evidence that the backdoor succeeded on more endpoint configurations.

What happened

On November 13, 2026, Cursor disclosed that a key exposed at Fireworks led to copied coding-agent traces, secondary credential access and the marketplace compromise. Cursor says fewer than 0.01% of customers had API credentials exposed and fewer than 0.05% of non-Privacy-Mode users had coding-agent traces exposed. The accessed data also included limited support information, archived non-production repositories, usage data and credentials previously sent to support.

Between 2:27 a.m. and 1:35 p.m. Pacific on October 29, 2026, malicious Anthropic.claude-code versions 2.1.312 and 2.1.313 were available through Cursor’s VS Code extension marketplace. Cursor says at most 5,000 users may have received the malicious builds during an approximately 11-hour exposure window. The affected agent component was the Anthropic.claude-code extension distributed through Cursor’s VS Code extension marketplace. The cited sources do not identify the underlying model or model version used by the malicious extension copies. The attacker modified the extension and used a stolen marketplace credential to place the copies in Cursor’s cache. The malicious code collected host metadata, added an SSH key labelled svc-telemetry, attempted to enable SSH, opened a reverse tunnel and installed persistence.

Published indicators include 134[.]209[.]82[.]225, ~/.config/.smt, ~/.config/.smtk, com.apple.sysmond.plist and the Windows scheduled task SysMon. The modified VSIX SHA-256 values are fcbb327e38e61c4fec795e0d0591f52cd1014b813d20a5912104953c3a17c364, 3de1722901165769893bef57c72594e9a54459faa23f3091b136b79966edd081, 80402fdcad8799285fd93044976973ca000da7995fb93363f819bb8e4f84ee6c and 6786ec90cbd45949695bd7805c9fbb26cfc91bc5fc897de0183d9bcb9cb21731. The published payload SHA-256 is ecb58cb29078d9646ad1dcdbe51526f2d9cfea06be5fb2ab183e8ecc3ab6bf15. Cursor says the last attacker activity it observed was on November 1, 2026, and that activity was blocked. Attribution posture: Cursor names no actor for the marketplace compromise; Fireworks separately describes links in its broader October incident to a financially motivated cybercriminal group. The cited source did not publish the underlying model detail described as Underlying model and version.

Why this matters now

The incident joined three trust layers that enterprises often assess separately: an AI infrastructure provider, historical coding-agent traces and a software-extension marketplace. A key exposed at one provider enabled access to data containing other credentials, which then enabled tampering with software delivered to developer endpoints.

Developer workstations commonly hold source-control tokens, cloud credentials, package-publishing rights and access to production support systems. Even though coding errors limited the backdoor on many hosts, an affected endpoint cannot be closed by uninstalling the extension alone because the payload attempted to leave independent SSH keys and persistence.

The disclosure also changes the governance question for AI coding tools. Privacy settings, training-data sharing, support submissions and historical traces can create durable secret stores. Security leaders need evidence that credentials are redacted before storage and that vendor keys cannot cross customer, analytics and distribution boundaries.

The decision for security leaders

Require endpoint and developer-platform owners to identify installations by extension version, marketplace source and installation time. Absence of the extension today does not prove the payload never executed or left persistence.

Treat positive endpoint indicators as a credential-containment event. Isolate the host, preserve evidence, remove the SSH key and persistence, assess rebuild requirements and rotate credentials from a trusted system.

Review AI-provider data flows as a security architecture issue. Document which traces, prompts, support records and secrets leave the enterprise; enforce redaction, retention limits, least-privilege vendor keys and provenance controls for extension distribution.

Evidence of closure

  • Extension inventory identifies every installation by version, source and time.
  • Endpoint hunt shows no published artefacts or records a rebuild for each positive host.
  • Credential register proves all secrets reachable from affected hosts were rotated.
  • AI-tool data-flow review documents approved retention, redaction and Privacy Mode settings.

The Security.io assessment

Cursor’s expanded disclosure materially strengthens the evidence chain from an exposed third-party key to stored traces, additional credentials and marketplace tampering. It also narrows scope: Cursor says official Anthropic distribution, Open VSX, customer repositories, Privacy Mode data and its production release-signing path were not affected.

The backdoor’s coding errors may have prevented successful remote access on many Windows and macOS hosts, particularly where enabling SSH required privileges the extension lacked. That limitation does not close affected endpoints because host details, an SSH key and persistence scripts could still have been written.

Vendor containment is not enterprise closure. Each organisation must establish whether affected builds reached its devices, whether payload artefacts exist and which credentials were accessible. Fireworks’ separate bulletin also reinforces the need to scope vendor-held secrets beyond the immediate Cursor incident.

Questions for the morning meeting

  • Which developers received the affected extension versions during the exposure window?
  • What production credentials were accessible from potentially affected workstations?
  • Which AI vendors receive historical coding-agent traces or support data?
  • Can extension marketplaces prove package provenance before local installation?

Related intelligence

Shared decision context