A generated abstract illustration on a dark navy field: a thin cyan line of small angular glyph shapes runs from a weathered grey rectangle at the left, passes through a translucent amber pane marked only by a faint grid, and terminates at a glowing white terminal outline at the right. Three small pale file tiles are pinned along the line like waypoints. No text, people or organisation logos appear. The OrcaRouter logo is composited in the bottom-right corner.
AI Safety Incidents

IDEsaster: 30+ Flaws, 24 CVEs, and One Injection Pattern That Reached Every AI IDE

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On 6 December 2025 the security researcher Ari Marzouk, who publishes as MaccariTA, released IDEsaster, the write-up of a six-month investigation that found more than 30 vulnerabilities in AI-powered code editors and coding assistants, and concluded that 100% of the AI IDEs and assistants he tested could be driven from a poisoned file or a hostile Model Context Protocol server into data theft or remote code execution. Twenty-four of the flaws carry CVE identifiers, and an AWS security bulletin published two months earlier credits him among the researchers behind prompt-injection fixes in Amazon Q Developer and Kiro. The affected list is most of the agentic-coding market: GitHub Copilot, Cursor, Windsurf, Kiro.dev, Zed.dev, Roo Code, JetBrains Junie, Cline, G​emini CLI and C​laude Code. Nobody, including the researcher, has reported the chain being used against a real target. This is one record in the AI incident archive, where each entry carries its own sources, severity, confidence grade and disputed flag.

What the attack chain actually was

Marzouk's argument, laid out in the report and in an interview with The Hacker News, is not that any single assistant had a bug. It is that AI coding tools were built on top of software that had spent twenty years assuming a human was the only thing operating it, and that no vendor had moved the base editor into its threat model.

Earlier disclosures in the same space each ended at a single application: a tool with a path traversal, an agent configuration file left writable, an MCP configuration an agent could edit. IDEsaster adds a third step to the chain. First a prompt injection hijacks the agent's context. Then the agent's own tools — often working exactly as designed, with in-workspace file writes auto-approved by default — perform ordinary actions. Then a long-standing feature of the base IDE, triggered by whichever file the agent just wrote, leaks data or runs code. Because the last step belongs to the shared editor underneath, every assistant built on that editor inherits it.

The injection itself is the cheap part. Marzouk's threat model lists rule files, MCP servers that are compromised, rug-pulled or merely relaying attacker-controlled content, deeplinks, pasted URLs, file names, and text hidden in a README. The Hacker News summarised the same list. "All AI IDEs (and coding assistants that integrate with them) effectively ignore the base software (IDE) in their threat model," Marzouk told the outlet. "They treat their features as inherently safe because they've been there for years. However, once you add AI agents that can act autonomously, the same features can be weaponized into data exfiltration and RCE primitives."

Three ways it ended in data theft or code execution

Remote JSON schemas. A JSON file may name an external schema in a $schema field, and the editor fetches that URL to validate against it. Write a JSON file whose schema URL carries data in its query string, and the editor makes the request itself — a GET to the attacker's server with the stolen content attached. Marzouk reports that Visual Studio Code, JetBrains IDEs and Zed all support the feature by default, and that the request fires even when a diff preview is shown, which may bypass the human approval step some tools rely on.

IDE settings that name an executable. Every Git repository ships sample hooks under .git/hooks/, and they are writable. An agent that edits one of those files can then write a line into .vscode/settings.json pointing php.validate.executablePath at it, and create any PHP file. The editor runs the configured "PHP binary" the moment it sees the file. The JetBrains variant uses PATH_TO_GIT inside .idea/workspace.xml instead.

Multi-root workspaces. Visual Studio Code's *.code-workspace file can also add a folder outside the project as a workspace root, which defeats the out-of-workspace approval prompt. The attacker then edits an executable that is now inside the workspace, points php.validate.executablePath at it, and triggers it. Marzouk calls this third case the one that shows the real risk: block one settings file and the next feature offers the same primitive.

Marzouk withheld the exact exploitation prompts in all three cases, writing that some vendors had acknowledged the issues but not fixed them despite more than 90 days of responsible disclosure. The report gives the affected products and the CVE numbers instead, which is enough to check the claims against the public records. The identifiers, the severity scores and the fixed versions all reconcile: Cursor's JSON-schema leak is CVE-2025-49150 at 5.9, fixed in 0.51.0; its settings-file path is CVE-2025-54130 at 7.5, fixed in 1.3.9; Roo Code's equivalents are CVE-2025-53097 at 5.9 and CVE-2025-53536 at 8.1. Zed's agent panel is CVE-2025-55012, rated 8.5 under CVSS v4.0 and fixed in 0.197.3. Microsoft's two Copilot entries, CVE-2025-53773 and CVE-2025-64660, are rated 7.8 and 8.0, and MSRC records both as Important, not publicly disclosed and not exploited.

An HTML card headed 'Ten of the 24 CVEs, and what each one actually fixed' with the kicker 'IDEsaster · published 6 December 2025'. A ten-row table lists CVE identifier, product, the feature abused, the CVSS score and the fixed version: CVE-2025-49150 Cursor, remote JSON schema fetched on write, 5.9 MED, fixed in 0.51.0; CVE-2025-53097 Roo Code, search_files ignoring the workspace limit then schema, 5.9 MED, 3.20.3; CVE-2025-58335 JetBrains Junie, information disclosure via search_project, 5.5 MED, 252.284.66 and later; CVE-2025-54130 Cursor, agent writes .vscode/settings.json with no approval, 7.5 HIGH, 1.3.9; CVE-2025-53536 Roo Code, executable path in .vscode/settings.json, 8.1 HIGH, 3.22.6; CVE-2025-55012 Zed, agent bypasses permission checks for config, 8.5 HIGH under CVSS v4.0, 0.197.3; CVE-2025-53773 GitHub Copilot and Visual Studio, command injection into a settings file, 7.8 HIGH, Visual Studio 17.14.12; CVE-2025-64660 GitHub Copilot and Visual Studio Code, improper access control in a multi-root workspace, 8.0 HIGH, Visual Studio Code 1.106.2; CVE-2025-61590 Cursor, RCE via .code-workspace files, 7.5 HIGH, 1.7; CVE-2025-58372 Roo Code, RCE via .code-workspace files, 8.1 HIGH, 3.26.0. A footer notes that CVE-2025-58372 was reported to Roo Code by thelicato rather than by Marzouk, and that the report counts 30+ vulnerabilities in total so several findings carry no identifier. The OrcaRouter logo is composited in the bottom-right corner.

Who was affected, and how the vendors answered

The exposed population is anyone running an agentic assistant inside Visual Studio Code, a JetBrains IDE or Zed, which between them carry most of the commercial and open-source coding agents in use. Marzouk describes the products as "10+ market-leading products affecting millions of users"; that figure is his characterisation of reach, not a measured count of installations, and no source publishes one. Tom's Hardware, reporting the study on 7 December 2025, repeats the 100% finding and the ten named products.

The vendor responses were uneven, and the gaps matter more than the patches. GitHub Copilot's and Kiro's remote-schema issues were both fixed without a CVE being assigned, and so was Kiro's settings-file path, according to the report; Copilot's settings and multi-root cases did get identifiers. JetBrains shipped builds of Junie with the fix. Cursor, Roo Code, Zed and Microsoft issued numbered CVEs with fixed versions a defender can check. C​laude Code, according to the report, acknowledged the issue and chose to address it with a security warning in its documentation rather than a code change — the C​laude Code security documentation does now carry a trust section, workspace-boundary prompts and MCP approval guidance, though it does not name IDEsaster.

The one advisory outside the CVE system is AWS-2025-019, published by Amazon on 7 October 2025, which covers Amazon Q Developer and Kiro and credits Embrace the Red, HiddenLayer and MaccariTA. It records the fixes as they landed: Language Server 1.22.0 on 17 July 2025 requiring human confirmation for find, grep and echo; 1.24.0 on 29 July for ping and dig; and Kiro 0.1.42 on 1 August requiring confirmation in Supervised mode. That bulletin is dated two months before IDEsaster was published, which is the clearest evidence that the underlying issues were being worked through vendors one at a time before the unifying write-up existed.

A generated abstract illustration on a dark navy field: at the left, a small pale document icon sits inside a dashed rectangle labelled only by shape, representing a project folder. A thin amber line runs from it into a tall dark slab of stacked horizontal bars at right, and at the point of contact a single bright white square opens like a shutter. Behind the slab, faint ghosted outlines of many more identical slabs recede into the dark. No text, people or organisation logos appear. The OrcaRouter logo is composited in the bottom-right corner.

What is established, and what is not

The archive grades this record confidence A, not disputed, with no real harm recorded, and the public record supports that. The CVE identifiers, their scores and their fixed versions are checkable in the numbering authority's own database and in Microsoft's advisory service, and they match the report. Tom's Hardware covered the study independently of The Hacker News. ReversingLabs published its own analysis on 25 March 2026, treating IDEsaster as a supply-chain risk category rather than a single incident. The AWS bulletin independently confirms the Kiro and Amazon Q strand and names Marzouk among the disclosing researchers.

Established, on the record:

• The mechanism, in three documented variants, with the specific IDE features named: remote JSON schema fetching, executable paths in settings files, and multi-root workspace roots.

• The identifiers. Twenty-four CVEs, ten of which are set out in the table above with scores and fixed versions taken from their published records, and AWS-2025-019 as the separate advisory.

• The fixes. Each CVE names a version, and the vendors shipped them between June and November 2025.

• That no exploitation has been reported. ReversingLabs' March 2026 piece quotes KnowBe4's Roger Grimes saying he was "unaware of a single real-world attack involving both AI and an IDE" — three months after publication, and consistent with the archive's no-real-harm flag.

Not established, or stated less precisely than it sounds:

• "Millions of users." That is the researcher's estimate of reach, and it is plausible given the products involved, but no vendor or regulator has published an affected-installation count. The archive's medium severity is a classification of the finding, not a measure of blast radius, and it sits below the 7.5-to-8.5 scores several of the individual CVEs carry.

• Whether every vendor finished. Marzouk wrote at publication that some had acknowledged the issues without fixing them after more than 90 days. He did not name them in the published report, so the claim cannot be checked vendor by vendor from the public record.

• Attribution of the work. The report does not state Marzouk's employer. ReversingLabs, writing in March 2026, describes him as a Microsoft senior security researcher; nothing in the primary documents reviewed for this piece confirms or denies that, and it does not change the technical findings.

• Whether 24 is the whole count. The report says "over 30" vulnerabilities and 24 CVEs, which leaves several findings without an identifier — the same gap that lets a fixed bug go unflagged in a vulnerability feed.

An HTML card titled 'Established, and not' with the kicker 'Record 2025-12-06-idesaster', split into two columns under the line that the archive files this as a research finding with confidence A, not disputed, no real harm recorded and date precise to the day. The left column, 'Established, on the record', lists five ticked items: 30+ findings and 24 CVEs published by Ari Marzouk on 6 December 2025 after a six-month investigation; three abuse variants against named editors; fixed versions named for every CVE in the table; independent corroboration from Tom's Hardware and from AWS-2025-019 crediting Marzouk on Kiro and Amazon Q; and no real-world use, with ReversingLabs in March 2026 quoting KnowBe4's Roger Grimes saying he knew of none. The right column, 'Not established, or looser than it sounds', lists five dashed items: the number of affected users; which vendors left issues unfixed at publication, whom the report does not name; the researcher's employer, which the report does not state; whether every one of the 30+ findings has a CVE; and the archive's medium severity, a classification rather than a blast-radius measure. A footer cites the IDEsaster report, The Hacker News, Tom's Hardware, AWS-2025-019 and ReversingLabs. The OrcaRouter logo is composited in the bottom-right corner.

What changed afterwards

The most concrete consequence is a second round of research. On 11 May 2026 Marzouk published IDEsaster 2.0, arguing that the mitigations vendors shipped in response — denylisting .vscode/settings.json, *.code-workspace and .idea/workspace.xml so an agent must ask before writing them — are structurally insufficient. The follow-up moves the gadget one layer outward, into legitimate language servers: an empty Gemfile.lock plus a crafted ruby_lsp/**/addon.rb is loaded and activated by Shopify's Ruby Language Server, which the report gives as having more than 30 million downloads, and a Directory.Build.props is evaluated by the C# language server, listed at more than 50 million. Both were validated as code execution paths. In that post he estimates that nearly 30 vulnerabilities were patched across 12 vendors after the first publication, alongside sandboxing and egress controls. Those figures are the researcher's own accounting, not an audited count.

The mitigations he recommends have not changed in substance. For people using these tools: work only in trusted projects, treat rule files, READMEs and even file names as injection vectors, connect only to MCP servers you monitor for changes, inspect pasted sources for hidden instructions, and turn human-in-the-loop back on wherever it exists. For people building them: scope each tool to the narrowest resource set, require approval for writes to dotfiles and configuration, sandbox command execution, allowlist egress at the editor layer, and assume the agent is already breached.

The durable point is narrower than the headline count. Every one of these bugs was a feature working as specified, in software that had no reason to expect an autonomous process to be the one calling it. That is why the fixes are scattered across a dozen vendors and two rounds of research rather than collected in one patch — and why the record is filed as a research finding rather than an incident.