Cursor Sovereign Runtime: The Editor That Writes Code Needs a Gate on Every Line

Cursor has the most complex CVE record of any platform in this series: 11-plus disclosed vulnerabilities, one currently unpatched, and four distinct architectural attack patterns. Every other runtime in the program had one or two novel architectural risks. Cursor has four, and the release names each one, gives its CVE or case name, and states exactly what closes it. The Cursor Sovereign Runtime, released July 8, 2026, reports 21 of 21 tests passing on the first pass, the first time in the series a runtime passed on its first run.

The problem: four patterns, and one of them has no patch

The first pattern is the most documented. The release calls it Pattern 2, indirect prompt injection to file write to remote code execution, and the case is CurXecute, CVSS 8.6. An indirect prompt injection writes to .cursor/mcp.json, and Cursor executes it immediately. The injection does not need to persuade the agent to run anything. It writes a configuration file the IDE trusts, and the IDE runs what the file says. The release states it is closed by two scans working together: scan_mcp_json() and scan_repo_file().

The second pattern is the most insidious, and the release’s word is exact. The case is MCPoison, CVSS 7.2, and the mechanism is a trust-on-first-use bypass. Approve-once trust binds to the key name, not the command body, so an attacker swaps the command after approval and gets silent execution on every IDE launch, forever. The approval was real. The command it approved is not the command that runs. The release states it is closed by scan_mcp_server_registration() with command-hash comparison, which binds the approval to the exact command that was approved, so a swapped command no longer matches.

The third pattern is the highest CVSS in the series. The case is CVE-2026-26268, scored 9.9 Critical, and the mechanism is a sandbox escape through a side channel. A sandboxed agent writes to .git/config, and a git hook executes outside the sandbox on the next git operation. The release names a second instance of the same pattern, NomShub, which uses the same technique through .zshenv. The sandbox holds the agent. It does not hold the files the agent writes, and the next tool that reads those files runs outside the sandbox. The release states it is closed by scan_repo_file() together with scan_shell_command().

The fourth pattern is the one with no fix. TrustFall is unpatched, according to the release. Project-local MCP servers execute without approval. There is no code fix, so the release gives a configuration mitigation instead: set "cursor.mcp.disableProjectLocalServers": true in managed workspace config, and enforce it at the config layer until Anysphere ships a patch. That is the honest shape of a runtime release: three patterns closed by code, and one pattern held at the configuration layer because the vendor has not patched it. A runtime that pretends the fourth pattern is closed would be lying. This one says plainly that it is not closed, and tells you what to do instead.

What the release changes

The runtime covers eight surfaces, the most of any platform in the series, and enforces 18 controls from AI SAFE2 v3.0, with a series total the release puts at more than 20 verified control IDs across five pillars plus the Cross-Pillar layer. The test result is the first clean first pass in the program: 21 of 21 on the first run. The release credits the test-first methodology hardened by the four prior builds, which produced cleaner patterns the first time. That is a claim about the program’s own learning, and it is consistent with the record: the Lovable release caught three gaps in testing before shipping, and the Langflow release caught a severity misclassification before shipping. By the fifth build, the patterns were clean enough to pass first time.

The runtime also plugs into the NEXUS mesh. The AISAFE2Engine kernel is shared across the runtimes in the program, so one engine instance gives one unified Love Score and one SHA-256 audit chain across the whole agentic stack, according to the release. One engine watches every runtime the organization runs, and the same audit chain covers all of them.

What to do with it

Get the package from examples/cursor-sovereign-runtime in the ai-safe2-framework repository. Then do the configuration mitigation first, before anything else, because TrustFall is unpatched and project-local MCP servers execute without approval until the setting is enforced. Set "cursor.mcp.disableProjectLocalServers": true in managed workspace config. Then run the smoke test the release documents, and test the three closable patterns yourself: an injection that tries to write .cursor/mcp.json, a command swapped after approval, and a sandboxed write to .git/config or .zshenv. A passing test is evidence for the case tested. It is not a general safety claim, and it should not be recorded as one.

Where it sits in the program

This release came on July 8, the same day as the Grok v3.1 update, in the middle of the July sequence. The sovereign runtime program began with the Hermes runtime in May 2026, followed by Anti-Gravity v2 on June 9. LangChain came next on June 19, then LangGraph on June 22, Langflow on June 24, AutoGen on June 26, CrewAI and Make.com on June 30, Grok on July 2 with a v3.1 update on July 8, Cursor on July 8, and Lovable on July 19. Cursor matters in that sequence because it was the hardest platform in the program to defend, with the most vulnerabilities, the most attack patterns, and the one unpatched flaw, and it was also the first runtime to pass its tests on the first attempt. Difficulty and discipline arrived together.

The controls, in one place

The runtime enforces 18 controls from AI SAFE2 v3.0: P1.T1.2, P1.T1.9, P1.T1.10, P1.T1.4_ADV, P1.T2.1, P1.T2.5, P1.T2.6, S1.3, S1.5, S1.6, P2.T3.1, A2.5, F3.2, P4.T7.1, M4.5, E5.1, CP.9, and CP.10. The release states that all controls were verified against the live AI SAFE2 v3.0 framework, with zero fabricated IDs. If a control ID in a future version does not map, that is the first thing to check.

What this does not establish

The 21 of 21 test result is the release’s own smoke test, and the release body does not describe any independent validation. Treat the number as the release’s claim about its own tests, not as outside proof. The CVE numbers and CVSS scores in this draft are the release body’s own figures, and they were not independently re-verified for this draft. Most important, TrustFall is unpatched as of this release, and the draft does not claim otherwise. Whether a patch has shipped since July 8, 2026 is not established in the sources for this draft. The configuration mitigation is a hold, not a fix, and it should be recorded as one until a patch is confirmed.

Sources

  • Cursor Sovereign Runtime release, ai-safe2-framework, tag 2026-07-08_Cursor_Sovereign_Runtime, published 2026-07-08: https://github.com/CyberStrategyInstitute/ai-safe2-framework/releases/tag/2026-07-08_Cursor_Sovereign_Runtime
  • Cursor Sovereign Runtime package: https://github.com/CyberStrategyInstitute/ai-safe2-framework/tree/main/examples/cursor-sovereign-runtime
  • AI SAFE2 framework repository: https://github.com/CyberStrategyInstitute/ai-safe2-framework
  • Release-notes audit, 2026-11-17: ~/workspace/seo-review/release-notes-audit-2026-11-17.md

Recent Posts

Tag Cloud

Stop Threats Before They Execute

Your free Kernel-Level Defense Buyer’s Guide is ready to download.

By providing my email address, I consent to receive emails and text messages—including newsletters and marketing communications—from creators of Warden Secure, Cyber Strategy Institute, our flagship zero-trust platform for ransomware prevention, and agree to the Terms and Privacy Policy. You may unsubscribe at any time.