AI Threat Alert
CVE-2026-87983: Mistral Vibe: quoted paths bypass file read guard
CRITICALMistral Vibe, an AI coding agent, is supposed to confine the shell commands it runs on a developer's behalf to the active project workspace, but a flaw in how it parses quotation marks lets an attacker slip in a quoted absolute path — for example a quoted "/home/user/.ssh/id_rsa" — that the sandbox check fails to catch, resulting in silent, unapproved reads of any file on the host. The vendor rates this critical, and the risk is amplified because the agent's own allowlisted shell commands, normally a trust boundary, become the exploitation vector, so no separate code-execution bug is needed, just a crafted path string reachable via user input, a repo file, or an untrusted prompt. There is no confirmed CISA KEV listing, no public exploit code, and no CVSS/EPSS score published yet, so mass opportunistic exploitation is not demonstrated, but a sandbox/path-validation bypass in a widely distributed agentic coding tool (472 downstream dependents, 8 prior CVEs in this package) is exactly the primitive that gets weaponized fast once a PoC circulates. Until Mistral ships a patched release, disable or tightly scope the affected shell-command allowlist, run Vibe under an OS-level sandbox or restricted service account with no access to SSH keys, cloud credentials, or .env files, and monitor agent-invoked shell command logs for quoted or absolute paths pointing outside the workspace directory.
What is the risk?
Rated critical by the vendor and independently plausible given the finding: this is a sandbox-escape class bug in a tool whose entire security model depends on confining file/command access to a workspace boundary. No CVSS vector, EPSS score, KEV listing, public exploit, or Nuclei template exists yet, so near-term mass exploitation is unconfirmed — but the bar to exploit is low (a malformed quoted path string, no memory corruption or auth bypass chain required) and the package has a track record of 8 prior CVEs, suggesting recurring hardening gaps. The package risk score (32/100) understates urgency here because it doesn't yet reflect this disclosure. Treat as high-priority-patch-now despite the absence of active-exploitation telemetry, because the exposure (arbitrary local file read, including credential material) is severe and the exploit primitive is trivial to reproduce once public.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| Mistral AI | pip | — | No patch |
Do you use Mistral AI? You're affected.
How severe is it?
What should I do?
1 step-
1) Check for and apply a patched Mistral Vibe release as soon as it is published (currently no fixed version listed) — track the vendor advisory at hiddenlayer.com. 2) Until patched, do not rely on Vibe's internal workspace restriction as a security boundary: run the agent inside a container, VM, or restricted OS user account with no access to SSH keys, cloud credential files, or unrelated repositories. 3) Where possible, disable or restrict the allowlisted shell commands that accept file paths, or wrap them with an external validator that normalizes and rejects quoted/escaped absolute paths independent of Vibe's own check. 4) Add detection: log and alert on Vibe-invoked shell commands containing quoted absolute paths or path traversal sequences (
../,~, leading/) resolving outside the active workspace root. 5) Rotate any credentials (SSH keys, cloud tokens, API secrets) that were reachable by a Vibe agent process before this advisory, if exploitation cannot be ruled out.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-87983?
Mistral Vibe, an AI coding agent, is supposed to confine the shell commands it runs on a developer's behalf to the active project workspace, but a flaw in how it parses quotation marks lets an attacker slip in a quoted absolute path — for example a quoted "/home/user/.ssh/id_rsa" — that the sandbox check fails to catch, resulting in silent, unapproved reads of any file on the host. The vendor rates this critical, and the risk is amplified because the agent's own allowlisted shell commands, normally a trust boundary, become the exploitation vector, so no separate code-execution bug is needed, just a crafted path string reachable via user input, a repo file, or an untrusted prompt. There is no confirmed CISA KEV listing, no public exploit code, and no CVSS/EPSS score published yet, so mass opportunistic exploitation is not demonstrated, but a sandbox/path-validation bypass in a widely distributed agentic coding tool (472 downstream dependents, 8 prior CVEs in this package) is exactly the primitive that gets weaponized fast once a PoC circulates. Until Mistral ships a patched release, disable or tightly scope the affected shell-command allowlist, run Vibe under an OS-level sandbox or restricted service account with no access to SSH keys, cloud credentials, or .env files, and monitor agent-invoked shell command logs for quoted or absolute paths pointing outside the workspace directory.
Is CVE-2026-87983 actively exploited?
No confirmed active exploitation of CVE-2026-87983 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-87983?
1) Check for and apply a patched Mistral Vibe release as soon as it is published (currently no fixed version listed) — track the vendor advisory at hiddenlayer.com. 2) Until patched, do not rely on Vibe's internal workspace restriction as a security boundary: run the agent inside a container, VM, or restricted OS user account with no access to SSH keys, cloud credential files, or unrelated repositories. 3) Where possible, disable or restrict the allowlisted shell commands that accept file paths, or wrap them with an external validator that normalizes and rejects quoted/escaped absolute paths independent of Vibe's own check. 4) Add detection: log and alert on Vibe-invoked shell commands containing quoted absolute paths or path traversal sequences (`../`, `~`, leading `/`) resolving outside the active workspace root. 5) Rotate any credentials (SSH keys, cloud tokens, API secrets) that were reachable by a Vibe agent process before this advisory, if exploitation cannot be ruled out.
What systems are affected by CVE-2026-87983?
This vulnerability affects the following AI/ML architecture patterns: agent frameworks, coding agents / developer tooling.
What is the CVSS score for CVE-2026-87983?
No CVSS score has been assigned yet.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0037 Data from Local System AML.T0050 Command and Scripting Interpreter AML.T0053 AI Agent Tool Invocation Compliance Controls Affected
What are the technical details?
Original Advisory
An arbitrary file read vulnerability in Mistral Vibe, introduced in version 2.6.0, allows an attacker to bypass workspace restrictions using quoted absolute paths in allowlisted shell commands. Improper handling of quotation marks during path validation enables files outside the active workspace to be read without user approval.
Exploitation Scenario
An attacker plants a malicious instruction — via a poisoned file in a public repository the victim opens with Vibe, a crafted issue/PR description, or a direct prompt if the attacker has any interactive access — that induces the agent to invoke one of its allowlisted shell commands with a quoted absolute path, such as a quoted path to `~/.ssh/id_rsa`, `~/.aws/credentials`, or a `.env` file outside the project. Because the path-validation logic mishandles quotation marks, the workspace-restriction check is bypassed and the command executes without prompting the user for approval. The file contents are then returned into the agent's output/context, where the attacker (via prompt injection) can have the agent echo, summarize, or otherwise relay the secret material back to them, completing a credential-theft chain with no traditional code execution required.
Weaknesses (CWE)
CWE-22 — Improper Limitation of a Pathname to a Restricted Directory ('Path Traversal'): The product uses external input to construct a pathname that is intended to identify a file or directory that is located underneath a restricted parent directory, but the product does not properly neutralize special elements within the pathname that can cause the pathname to resolve to a location that is outside of the restricted directory.
- [Implementation] Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylis
- [Architecture and Design] For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.
Source: MITRE CWE corpus.
Timeline
Related Vulnerabilities
GHSA-wx9m-wx4f-4cmg 9.6 mistralai 2.4.6: supply chain dropper executes on import
Same package: mistralai CVE-2026-67623 8.8 Mistral Vibe: RCE via malicious git fsmonitor hook
Same package: mistralai CVE-2026-93993 8.8 Analysis pending
Same package: mistralai CVE-2026-87984 Mistral Vibe: shell redirection bypasses file-write guardrails
Same package: mistralai GHSA-jgg6-4rpr-wfh7 Mistral npm SDK: supply chain attack, no impact
Same package: mistralai