AI in Insurance · Security notice

Cursor says tampered Claude Code extension in its marketplace tried to plant backdoors on developers' machines

Cursor says modified versions of Anthropic's Claude Code extension were served through its VS Code extension marketplace for about 11 hours on 29 October. The code collected host details and tried to open a persistent reverse SSH tunnel for the attacker. Ars Technica, citing Cursor, put affected users at no more than 5,000; neither company has explained how the attackers got in.

Illustration of parcels on a conveyor leading to an open laptop, with one parcel a different colour from the rest.
Illustration generated with AI for insurewire.org. Not a photograph of the event.

Cursor, which makes an AI-assisted code editor, says versions 2.1.312 and 2.1.313 of the Anthropic.claude-code extension in its VS Code extension marketplace were malicious and "were not developed by Anthropic". They were available from about 02:27 to 13:35 Pacific time on 29 October. Cursor's notice, updated on 30 October, names both versions; Ars Technica's report named 2.1.313. Ars reported that the tampered release went out through Anthropic's official publisher account and that, because Cursor updates extensions automatically by default, users could receive it without doing anything. Citing Cursor, Ars said no more than 5,000 users were affected.

According to Cursor, the code sent the host name, user name, operating system and public IP address to a server run by the attacker. It also tried to add the attacker's SSH key, switch on the local SSH service and open a reverse tunnel, and to keep that tunnel alive through a macOS launch agent or a Windows scheduled task that retried about every 60 seconds. Errors in the code meant the tunnel step would usually fail unless the editor ran with administrator rights. Security firm Phorion told Ars that the stolen names of environment variables would let the operator see which cloud credentials and API tokens sat on each machine and choose targets for follow-up attacks.

Cursor removed the versions, took its marketplace offline to rotate keys and audit access, and says no other extension was altered. It also worked with the hosting provider to take the attacker's server offline. It warns that uninstalling the extension may not be enough: affected users should isolate the machine, remove the key and persistence, rotate any credentials reachable from it and, if needed, rebuild it.

Neither Cursor nor Anthropic has said how the attackers gained control of the publishing channel. When Ars published on 30 October, Anthropic had not issued a public notice and neither company had answered its questions.

Read the source

Cursor ↗

Company · 29 Oct 2026

This is an original brief based on the linked publication. Company statements and forecasts are attributed to their source.

Continue reading

All stories →
Illustration

AI in Insurance

Dubai regulator publishes paper on how frontier and agentic AI could strain financial firms' cyber resilience

The Dubai Financial Services Authority and the Cambridge Digital Innovation and Regulation Initiative published an insights paper on 30 October on how frontier and agentic AI could change cyber resilience in financial services. It tests resilience on four dimensions: speed, reach, authority and trust. The DFSA says the paper is not guidance and creates no supervisory requirements.

Illustration

AI in Insurance

MS Amlin launches AI tool to check reinsurance wordings before the 1 January renewals

MS Amlin says an AI tool called Wordy Bot is in use on its North American property reinsurance book. The tool checks contract wordings against the Lloyd's insurer's underwriting standards and flags missing clauses and changes. MS Amlin expects it to halve wording review time before the 1 January renewals, but has given no measured result.

Illustration

AI in Insurance

Records obtained by EFF show CMS challenged four of six vendors in Medicare's AI prior-authorisation pilot

The Electronic Frontier Foundation has published more than 800 pages of CMS records on WISeR, the pilot that uses AI vendors to review prior-authorisation requests in traditional Medicare. They show CMS sent information requests or corrective-action demands to four of the six vendors in the first months, and a CMS document says the payment design gives vendors an incentive to deny claims.

The daily briefing