Science1 distinct publisher3 min readUpdated
Mindgard says opening a cloned project in Cursor on Windows silently executes any git.exe planted in the repo root. It went full disclosure after months of no response.
The Scientist · Science desk
Compiled by The ScientistSomething wrong?How this is made
Mindgard has published an unfixed flaw in Cursor on Windows: when the IDE loads a project it looks for git binaries in several locations, one of which is the current workspace, so a repository carrying a malicious `git.exe` in its root is executed automatically [1]. According to Mindgard there are no clicks, prompts, approval dialogs or warnings, and the result is arbitrary code execution [2]; the execution also recurs on a cadence rather than firing once [3].
That is the whole exploit. Mindgard notes it needs no prompt injection, no model manipulation, no jailbreak, no memory corruption and no sophisticated tradecraft, only a developer opening a project that contains the binary [4]. For anyone who clones repositories to read them, which is most of the job, the trust boundary has moved from "run the build" to "open the folder."
The blast radius is the reason to care. Mindgard cites Cursor's scale as more than 7 million active users, more than 1 million daily and more than 1 million paying, across more than 50,000 companies [5], against a reported market price of $60 billion [6].
The disclosure history is the part operators should read twice. Mindgard says it identified the issue on December 15, 2025 and reported it the same day, and multiple times since [7], and that more than six months and 197 or more new versions later the bug is still present in the latest version it tested [8]. Initial contact went to the security address in Cursor's published security.txt, followed by unanswered follow-ups and public attempts to find a contact [9]. Cursor's CISO eventually replied, attributing the gap to an internal automation failure that had stopped the expected HackerOne workflow, and invited Mindgard into the private bug bounty programme [10]. The resubmitted report was closed as Informative and out of scope; after a challenge, HackerOne reopened it, reproduced the issue and confirmed delivery to Cursor [11]. Mindgard says that after that, escalation through HackerOne and direct outreach to leadership both produced no response, while more than 70 versions shipped with the bug intact [12].
One caveat on the record: the post's own arithmetic does not reconcile. A December 15, 2025 discovery date sits awkwardly beside "more than six months," and the version counts of 197-plus and 70-plus appear in the same account without being squared [1]. We only have Mindgard's telling here; there is no Cursor statement in the material.
The interim advice is unglamorous. For managed Windows fleets, Mindgard suggests AppLocker or Windows App Control path-based deny rules scoped to workspace roots, such as `%USERPROFILE%\source\repos\*\filename.exe`, rather than hash rules, because attacker binaries vary by hash [13]. It also notes Windows has no general built-in way to block a child executable only when a specific parent launches it, so parent-aware enforcement needs EDR or a custom endpoint product [14]. For everyone else: open untrusted repositories in a VM, Windows Sandbox or another disposable environment, and do not rely on hash blocklists [15].
Watch for three things. Whether Cursor ships a release note that stops resolving git from workspace-relative paths, or at least prompts. Whether the HackerOne report changes state, given it has already been reopened and reproduced [11]. And whether your own endpoint tooling can see a git-named binary spawned from a repository root at all, because the path-based rule is a name filter and an attacker who renames the target loses nothing.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
After loading a project, Cursor attempts to find git binaries at various locations including the current workspace; a repository with a planted malicious git.exe in the root will be executed by the IDE.
A developer opening a repository in Cursor on Windows that contains a malicious git.exe in the project root triggers automatic execution with no clicks, prompts, approval dialogs or warnings; the result is arbitrary code execution.
The execution of the planted binary occurs repeatedly on a cadence.
The vulnerability does not depend on a complex exploitation chain, prompt injection, model manipulation, jailbreaks, memory corruption or sophisticated attacker tradecraft; exploitation requires only that a developer open a project containing a git.exe binary at the repository root.
Initial disclosure was sent to Cursor's security reporting e-mail address as specified in the company's published security.txt; follow-ups were sent when no confirmation was received, and public outreach was made to identify a security contact.
Cursor's CISO eventually responded and acknowledged that an internal automation failure had prevented the expected HackerOne workflow from taking place; Mindgard was invited into the private bug bounty program and resubmitted the report.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Detailed but single-source and vendor-unconfirmed
The technical core is unusually falsifiable for a single-publisher story: a named mechanism (workspace included in Git binary path resolution), a reproducible PoC (Calculator renamed git.exe at repo root), a Process Monitor process-create line showing Cursor.exe launching the repo-local binary for 'git rev-parse --show-toplevel', and a specific last-verified build and date (Cursor 3.2.16, April 30, 2026). Against that: one publisher, who is also the disclosing party; no CVE or tracking ID; no vendor or HackerOne statement; no independent reproduction; and internal timeline and version-count figures that do not reconcile.
Large stated install base, no measured exposure or remediation uptake
Exposure is plausibly wide because the affected product is a mainstream AI IDE, but the only numbers offered (7M+ active, 1M+ paying, 50K+ companies) are restated by the researcher without a primary citation, and they measure Cursor's user base rather than exposure to this issue: no figure is given for Windows share, for how many users open untrusted third-party repositories, for real-world exploitation, or for uptake of the suggested AppLocker/App Control mitigations. Continued releases through the disclosure window are asserted with two conflicting counts.
Concrete finding, modestly overstated framing
The finding itself looks understated in danger terms rather than inflated - silent, repeated execution of repository-supplied binaries needs no exploit chain. The overstatement is in the packaging: '0day' plus a $60 billion valuation and 7M-user framing for rhetorical weight, an implicit near-zero exploitation barrier that still requires the victim to clone and open a hostile repository, and elapsed-time and version counts inflated in one paragraph (six months / 197+ versions) relative to another (April 30 verification / 70+ versions). No vendor response, CVE or in-the-wild evidence is available to balance the account, so the gap is positive but small.
Disclosing security vendor publishing on a high-profile AI tool
The only publisher is Mindgard, a commercial AI-security research firm and the reporting party. Full disclosure of an unfixed flaw in one of the most visible AI coding tools serves its commercial visibility, and the article's editorial arc - 'what exactly is the security process for?' - is a positioning argument as much as a technical writeup, while also invoking the vendor's valuation and user counts. That incentive does not make the technical claim wrong; it does mean the process narrative, including the assertion that details reached Cursor and that remediation never started, is told solely by the party with an interest in its telling and is nowhere counterweighted.
Moderate: verifiable mechanism, unverified process and numbers
Confidence is held mid-range by a split: the technical claim is specific enough that a reader with Windows and Cursor could confirm or refute it quickly, and the mitigation advice stands on documented Windows capabilities, so those elements are dependable. Everything surrounding it - the disclosure chronology, remediation status, release counts, usage scale and valuation - depends on one interested publisher, is internally inconsistent in places, and has no vendor statement, CVE or second publisher to check against.
build
Waku 0.1.0 bets the product is the control plane, not another coding agent1 distinct publisher
build
Agent-written docs need a paper trail, not a confidence score1 distinct publisher
build
The kernel's Assisted-by trailer makes AI help a signed line, not a rumour1 distinct publisher
product
The allowlist read the command name, not what the command would do1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 19, 2026