All posts
Tools Published Updated 12 min

November 14 agent guide: Cursor security findings and Coding Agent Index v2.1

November 14, 2026 is a trust and measurement day for coding agents. Cursor published a full investigation update on the October marketplace incident, and Artificial Analysis refreshed its Coding Agent Index to v2.1 with harder FrontierCode grading.

Jigar JoshiJigar JoshiAgentic AI Architect and Consultant
November 14, 2026 cover for Cursor security findings and Coding Agent Index
In this post (8 sections)

November 14, 2026 continues the Bacancy Technology window with coding-agent trust and evals. Cursor published a detailed investigation update on the October 29 marketplace incident. Artificial Analysis shipped Coding Agent Index v2.1 with FrontierCode-AA 2.0 and Terminal-Bench 4.1, resetting how teams compare Claude Code, Devin, Gemini, and peers.

What builders should know

November 14 focus areas
ItemPrimary sourceStatusAction
Cursor incident investigationcursor.com/blog/security-incident-updateContained; last attacker activity blocked Nov 1Confirm Privacy Mode, rotate any notified keys, re-check extension provenance
Malicious Claude Code cache buildsCursor Oct 29 notice + Nov 13 update2.1.312 / 2.1.313 Windows and macOS, ~11 hoursFollow Cursor remediation if those versions were installed from Cursor’s marketplace
Coding Agent Index v2.1Artificial Analysis methodology322 tasks; harder FrontierCode gradingRe-baseline harness scores; do not compare to Index v2.0 numbers

Cursor’s investigation: how the attack unfolded

Cursor’s November 13 update reconstructs a multi-step chain. An access key Cursor had issued to Fireworks was exposed by Fireworks. The attacker used it to copy data from a read-only training bucket of coding-agent traces from sessions not in Privacy Mode (mostly 2025). A small fraction of those traces still contained credentials; Cursor reports under 0.01% of customers had API credentials exposed. From there the attacker reached selected SaaS tools, archived non-production source, limited support data, usage metrics, some conversation summaries, and credentials users had sent to support.

Marketplace cache malware

On October 29, using a stolen credential, the attacker placed unauthorized malicious copies of Anthropic’s Claude Code extension onto the cache serving Cursor’s VS Code extension marketplace. Cursor emphasizes Open VSX never carried the malware and Anthropic did not publish those builds. Exposure lasted about 11 hours. At most roughly 5,000 users who requested the specific cached versions may have received 2.1.312 or 2.1.313. Cursor removed the builds, took the marketplace offline, and worked with the hosting provider to shut down the attacker server.

What Cursor says was not affected

  • Prompts and code for Privacy Mode users.
  • Coding-agent traces for non-Privacy Mode users: exposed for fewer than 0.05% of users; Cursor says it notified those users.
  • Live traffic and current prompts; production Cursor infrastructure; release signing keys; customer repositories.
  • Other marketplace extensions after validation and rebuild; Cloud Agents, Grok Bot, and other SpaceX SI products as listed by Cursor.
  • Inference moved off Fireworks as a precaution.

Why this matters for agent platforms

The incident is a supply-chain lesson for every coding-agent vendor: training-data buckets, vendor keys, support ticket credentials, and extension caches are part of the agent attack surface. Bacancy Technology already covered Cursor Live Apps publish gates on November 5. This investigation raises a second control plane: marketplace integrity, Privacy Mode defaults, secret scanning in traces, and vendor key scoping. Pair it with November 10’s OpenShell CNCF move: enforce policy outside the model.

Artificial Analysis Coding Agent Index v2.1

Artificial Analysis’s Coding Agent Index v2.1 is the equal-weight average of FrontierCode-AA 2.0 (132 repository-change tasks graded with Cognition’s integrity scanner and LLM judge), Terminal-Bench 4.1 (66 terminal tasks with harder verifiers), and SWE-Atlas-QnA-AA (124 repository Q&A tasks). Each task runs three times. Version notes say Terminal-Bench moved from 4.0 to 4.1 and FrontierCode v1.1 was replaced by FrontierCode-AA 2.0, so Index scores are not continuous with prior editions.

What the harder bar does

FrontierCode 2.0 judges whether a maintainer would merge a patch. Cognition’s redesign adds more repositories, more blocking criteria, larger reference diffs, and thoroughness checks that fail incomplete sibling-path fixes. Secondary reporting on Artificial Analysis results notes mean FrontierCode scores falling on the order of 16 points across configurations that ran both editions, with overall Index scores down about five points. Top reported composite scores cluster near 58 for Claude Code with Claude Fable 5.2 at high or xhigh effort and Claude Opus 5.5 at max, with Devin Fusion CLI mixes also in that band. Cost per task among high scorers still spans a wide range. Raising effort to max can hurt scores when agents exceed scope or time limits.

Builder checklist

  1. 01
    Close the Cursor incident loop
    Confirm whether any machine installed the listed extension versions. Rotate notified keys. Enforce Privacy Mode for regulated codebases.
  2. 02
    Treat extension marketplaces as production dependencies
    Pin versions, prefer signed upstream channels, and inventory who can install IDE extensions in the enterprise image.
  3. 03
    Re-baseline coding-agent evals on Index v2.1 components
    Run FrontierCode-style merge criteria and Terminal-Bench 4.1-style checks in-house. Do not keep old Index targets.
  4. 04
    Separate effort knobs from quality gates
    Max reasoning is not always better. Cap time, file scope, and spend in harness configs used for vendor comparison.

Conclusion

November 14 is about whether coding agents can be trusted in the supply chain and scored under harder merge standards. Cursor’s investigation gives concrete blast-radius numbers and remediation. Artificial Analysis v2.1 resets leaderboard inflation. Bacancy Technology’s advice: Privacy Mode and marketplace controls first, then re-run evals before changing default agents.

Need coding-agent security and eval design?

Bacancy Technology helps security and platform teams harden agent IDEs, marketplace policy, and coding-agent harnesses. Consulting: /consulting/ . Contact: /contact/ .

Sources: Cursor changelog ; Cursor changelog ; artificialanalysis.ai — coding agents benchmarking ; artificialanalysis.ai — coding agents .

The weekly take

Agentic AI patterns, delivered Thursdays

What I am shipping, watching, and pruning out of client stacks each week. One email. No fluff.

Shipping an agentic AI project this quarter?
Book a 30-min consult
Frequently asked

Questions readers ask about this post

Share this post
LinkedIn Facebook WhatsApp