Openai
Monthly
Arbitrary code execution in OpenAI Codex CLI (Windows, macOS, Linux) and Codex Desktop (Windows, macOS) allows an attacker who can deliver a crafted repository archive to run code with the victim user's full privileges when Codex automatically collects Git repository metadata. The root cause is that Codex's Git metadata collection path did not suppress the repository-local core.fsmonitor configuration, enabling a malicious .git/config to direct Git to invoke an attacker-supplied filesystem-monitor helper outside Codex's command sandbox and without any user-approval prompt. OpenAI has issued a source-level fix via PR #22652 adding -c core.fsmonitor=false to the relevant Git invocations; no public exploit code has been identified at time of analysis.
Arbitrary code execution in OpenAI Codex Desktop (Windows and macOS, including the Microsoft Store package) occurs when a user opens a workspace containing an attacker-prepared Git repository whose local .git/config is intact. Because Codex automatically inspects Git metadata and working-tree status on open, a malicious attr.tree setting combined with a clean/process filter causes Git to launch an attacker-controlled program outside Codex's command sandbox, with the signed-in user's privileges and no trust prompt, approval, or model interaction. No public exploit is identified at time of analysis and EPSS is low (0.14%), but the outcome is full local compromise of the user's files and credentials.
OpenAI Codex CLI and Codex Desktop's PowerShell command-safety parser failed to recognize PowerShell's stop-parsing token (--%), misclassifying file-writing Git commands embedded after that token as safe and executing them without user approval. An attacker who tricks a developer into opening a crafted repository can cause Codex to silently write an attacker-controlled configuration; when Codex subsequently loads that configuration, it launches an attacker-designated MCP server and executes arbitrary code under the user's account. No public exploit has been identified at time of analysis (EPSS 0.16%, 6th percentile), and the default filesystem sandbox on macOS and Linux can limit the damage radius by preventing writes outside permitted directories.
Arbitrary code execution in OpenAI Codex Desktop for Windows and macOS allows an attacker's Git hooks to run outside the application's command sandbox when a user opens a specially prepared repository. Codex's automated Git operations honored the repository-local `core.hooksPath` setting in `.git/config` without sanitizing or overriding it, enabling a malicious hook script to execute with the user's full OS privileges - reading, modifying, or deleting files and accessing other resources. No active exploitation has been identified (EPSS 0.13%, not in CISA KEV); a code-level fix exists in GitHub PR #22843 but a versioned patch release is not independently confirmed at time of analysis.
Remote image auto-fetch in the OpenAI Codex desktop app for macOS (versions prior to 26.527.31326) enables silent exfiltration of session secrets via indirect prompt injection. An attacker who can place malicious instructions into content processed by Codex - such as a tool result, API response, or file read during a session - can manipulate the model into generating a Markdown image tag whose URL encodes sensitive data; the app then automatically fetches that URL, transmitting API keys, source code, or tool-returned data to an attacker-controlled server with no additional user action. No active exploitation is confirmed (not listed in CISA KEV), no public proof-of-concept is identified, and EPSS sits at 0.16% (6th percentile), indicating low current exploitation probability despite the high-value target profile of affected users.
OpenAI Atlas before 1.2025.288.15 improperly exposed privileged browser APIs - including browser history access and tab open/close controls - to any web content loaded from *.openai.com origins, including the publicly accessible forum.openai.com. Because forum.openai.com harbored an independent cross-site scripting vulnerability, an attacker could chain the two weaknesses: inject a script via the forum XSS, which then silently invokes Atlas's privileged APIs against a visiting victim. No public exploit or CISA KEV listing has been confirmed at time of analysis, though a public security research blog (hacktron.ai) describes the attack surface; the EPSS score of 0.02% (4th percentile) reflects low observed exploitation probability.
Fullscreen API Spoofing and UI Redressing in the handling of Fullscreen API and UI rendering in OpenAI Operator SaaS on Web allows a remote attacker to capture sensitive user input (e.g., login credentials, email addresses) via displaying a deceptive fullscreen interface with overlaid fake browser controls and a distracting element (like a cookie consent screen) to obscure fullscreen notifications, tricking the user into interacting with the malicious site.
The ChatGPT system through 2025-03-30 performs inline rendering of SVG documents (instead of, for example, rendering them as text inside a code block), which enables HTML injection within most modern. Rated medium severity (CVSS 6.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available and no vendor patch available.
Arbitrary code execution in OpenAI Codex CLI (Windows, macOS, Linux) and Codex Desktop (Windows, macOS) allows an attacker who can deliver a crafted repository archive to run code with the victim user's full privileges when Codex automatically collects Git repository metadata. The root cause is that Codex's Git metadata collection path did not suppress the repository-local core.fsmonitor configuration, enabling a malicious .git/config to direct Git to invoke an attacker-supplied filesystem-monitor helper outside Codex's command sandbox and without any user-approval prompt. OpenAI has issued a source-level fix via PR #22652 adding -c core.fsmonitor=false to the relevant Git invocations; no public exploit code has been identified at time of analysis.
Arbitrary code execution in OpenAI Codex Desktop (Windows and macOS, including the Microsoft Store package) occurs when a user opens a workspace containing an attacker-prepared Git repository whose local .git/config is intact. Because Codex automatically inspects Git metadata and working-tree status on open, a malicious attr.tree setting combined with a clean/process filter causes Git to launch an attacker-controlled program outside Codex's command sandbox, with the signed-in user's privileges and no trust prompt, approval, or model interaction. No public exploit is identified at time of analysis and EPSS is low (0.14%), but the outcome is full local compromise of the user's files and credentials.
OpenAI Codex CLI and Codex Desktop's PowerShell command-safety parser failed to recognize PowerShell's stop-parsing token (--%), misclassifying file-writing Git commands embedded after that token as safe and executing them without user approval. An attacker who tricks a developer into opening a crafted repository can cause Codex to silently write an attacker-controlled configuration; when Codex subsequently loads that configuration, it launches an attacker-designated MCP server and executes arbitrary code under the user's account. No public exploit has been identified at time of analysis (EPSS 0.16%, 6th percentile), and the default filesystem sandbox on macOS and Linux can limit the damage radius by preventing writes outside permitted directories.
Arbitrary code execution in OpenAI Codex Desktop for Windows and macOS allows an attacker's Git hooks to run outside the application's command sandbox when a user opens a specially prepared repository. Codex's automated Git operations honored the repository-local `core.hooksPath` setting in `.git/config` without sanitizing or overriding it, enabling a malicious hook script to execute with the user's full OS privileges - reading, modifying, or deleting files and accessing other resources. No active exploitation has been identified (EPSS 0.13%, not in CISA KEV); a code-level fix exists in GitHub PR #22843 but a versioned patch release is not independently confirmed at time of analysis.
Remote image auto-fetch in the OpenAI Codex desktop app for macOS (versions prior to 26.527.31326) enables silent exfiltration of session secrets via indirect prompt injection. An attacker who can place malicious instructions into content processed by Codex - such as a tool result, API response, or file read during a session - can manipulate the model into generating a Markdown image tag whose URL encodes sensitive data; the app then automatically fetches that URL, transmitting API keys, source code, or tool-returned data to an attacker-controlled server with no additional user action. No active exploitation is confirmed (not listed in CISA KEV), no public proof-of-concept is identified, and EPSS sits at 0.16% (6th percentile), indicating low current exploitation probability despite the high-value target profile of affected users.
OpenAI Atlas before 1.2025.288.15 improperly exposed privileged browser APIs - including browser history access and tab open/close controls - to any web content loaded from *.openai.com origins, including the publicly accessible forum.openai.com. Because forum.openai.com harbored an independent cross-site scripting vulnerability, an attacker could chain the two weaknesses: inject a script via the forum XSS, which then silently invokes Atlas's privileged APIs against a visiting victim. No public exploit or CISA KEV listing has been confirmed at time of analysis, though a public security research blog (hacktron.ai) describes the attack surface; the EPSS score of 0.02% (4th percentile) reflects low observed exploitation probability.
Fullscreen API Spoofing and UI Redressing in the handling of Fullscreen API and UI rendering in OpenAI Operator SaaS on Web allows a remote attacker to capture sensitive user input (e.g., login credentials, email addresses) via displaying a deceptive fullscreen interface with overlaid fake browser controls and a distracting element (like a cookie consent screen) to obscure fullscreen notifications, tricking the user into interacting with the malicious site.
The ChatGPT system through 2025-03-30 performs inline rendering of SVG documents (instead of, for example, rendering them as text inside a code block), which enables HTML injection within most modern. Rated medium severity (CVSS 6.5), this vulnerability is remotely exploitable, no authentication required, low attack complexity. Public exploit code available and no vendor patch available.