Microsoft's agent-framework package on PyPI was compromised on November 17th, 2026. Versions 1.26.1 and 1.26.2 contained malicious code that installs Python startup hooks via .pth files, which execute silently when the interpreter initializes. The malicious code collects credentials related to Azure, GitHub Actions, and various CI/CD environments, including environment variables, configuration files, and cloud identity tokens. It actively requests fresh Azure managed-identity and GitHub Actions OIDC tokens. The stolen data is compressed, base64-encoded, and exfiltrated to a webhook.site endpoint. The attack targets credentials from multiple tools such as pip, npm, Docker, and Git.
A command injection vulnerability exists in compose-lint Action versions 0.30.0 and 0.31.0 when the sarif-file input points to a file in a subdirectory that cannot be created. This allows a pull request to introduce symlinks with directory names containing newline and command sequences, causing the Action to execute unintended workflow commands such as creating annotations or folding job logs. The issue is fixed in version 0.32.0. The vulnerability does not affect setups where sarif-file is at the repository root.
compose-lint version 0.32.0 contains an incomplete fix for a previous secret exposure vulnerability (GHSA-jf3h-8jvx-vcrg). When a run exceeds 20,000 findings, the process stops before masking secret values, causing secrets from a sibling .env file to be exposed in output formats including text logs, JSON, and SARIF reports. This can occur in CI workflows that write secrets into the .env file beside the Compose file and lint pull requests from untrusted authors. No fixed release was available at the time of disclosure, but version 0.33.0 addresses this issue.
compose-lint versions prior to 0.25.0 follow symlinks when reading project files, potentially exposing contents of files outside the project in its output. This can occur when a pull request commits a symlink pointing to sensitive files on the CI runner. The exposure includes quoted lines from non-YAML files and key names from env_file targets, visible in job logs and code scanning outputs. The issue is fixed starting from version 0.25.0, which restricts symlink resolution outside the project and removes quoted source lines from parse errors.
compose-lint versions from 0.2.0 up to but not including 0.32.0 automatically read a .compose-lint.yml policy file beside the Compose file. If this file is a symlink pointing outside the project directory, compose-lint follows it and reads the target file. When the target file contains invalid YAML, the parse error message includes a quoted snippet of the target file's content, exposing about 30 characters around the error line. This can leak small fragments of sensitive data, such as credentials, to pull request authors via job logs and code scanning uploads. The issue is fixed in version 0.32.0, which refuses to read symlinked policy files resolving outside the working directory and no longer quotes parse error lines.
compose-lint versions 0.22.0 through 0.31.0 can inadvertently disclose secret values from a CI-generated .env file when those secrets are referenced in the image field of a Docker Compose file. This occurs because compose-lint resolves ${VAR} references from the .env file and includes the resolved values in its findings, logs, JSON output, and SARIF reports. Although GitHub masks secrets in job logs, it does not mask them in uploaded JSON or SARIF artifacts or Code Scanning alert text, potentially exposing secrets in pull request workflows that have access to the .env file. The issue is fixed in version 0.32.0, which quotes the variable reference instead of the resolved secret value in outputs.
The compose-lint GitHub Action prior to version 0.31.0 outputs text that can be interpreted by the GitHub Actions runner as workflow commands. This allows a malicious Compose file in a pull request to inject commands that create fake error or warning annotations, fold logs to hide real verdicts, or mask strings in the log output. The issue arises because the Action does not neutralize the legacy command syntax '##[' in its text report output. The vulnerability was fixed in version 0.31.0 by escaping these sequences and using a stop-commands token to prevent command processing during output.
compose-lint versions from 0.22.0 up to but not including 0.31.0 improperly follow symlinks when reading the .env file, allowing a symlink outside the project directory to be read. This can expose environment variables from the CI runner, including secrets in certain configurations, via log outputs in JSON or SARIF formats. The issue is fixed in version 0.31.0 by enforcing containment checks on .env reads and refusing .env files that resolve outside the project directory.
The pypi-publish GitHub Action versions before 1.14.3 included the emSign root CA certificate via the certifi bundle, which has since been removed. This change affects the TLS trust store used by Sigstore for Trusted Publishing and Digital Attestations. The vulnerability is tracked due to the revocation of this CA in the chain that Sigstore relies on. A fix is available in version 1.14.3.
A new Mini Shai-Hulud worm variant named Trinitite was detected targeting the npm package @7nohe/openapi-react-query-codegen. The attacker exploited a GitHub workflow vulnerability allowing any user to trigger npm publish via pull-request comments, resulting in rapid publication of malicious package versions. The worm uses obfuscation techniques to execute even when scripts are disabled, steals credentials from multiple developer and cloud platforms, and exfiltrates data via GitHub commits. It establishes persistence through systemd services and includes a destructive token revoke mechanism that can wipe user directories. This campaign is linked to the TeamPCP threat actor and appeared shortly after arrests of related suspects in Australia.
Versions of wktk/conflibot before 1.2.1 are vulnerable to command injection via crafted pull request branch names when run on the pull_request_target event. This occurs because git commands are constructed using string interpolation and executed through a shell, allowing attacker-controlled branch names with shell metacharacters to execute arbitrary commands. The vulnerability allows attackers to execute commands with access to repository secrets and a write-scoped GITHUB_TOKEN, potentially leading to secret exfiltration and unauthorized repository modifications. The issue is fixed in versions 1.2.1 and later.