CVE-2026-57142
PraisonAI is a multi-agent teams system. From version 4.5.87 to before version 4.6.61, PraisonAI recipe execution has a dangerous-tool polic
CVSS
7.8
High
EPSS
—
KEV
—
Exploit Today
—
0-100
Published: Oct 23, 2026 · Last modified: Oct 23, 2026 · CWE-78 · CWE-94 · CWE-863
Not enough EPSS history yet.
PraisonAI is a multi-agent teams system. From version 4.5.87 to before version 4.6.61, PraisonAI recipe execution has a dangerous-tool policy that is supposed to block default-denied tools unless the caller explicitly passes allow_dangerous_tools=True. That policy only checks tools declared in TEMPLATE.yaml requires.tools. For steps-based recipes, the actual execution path loads workflow.yaml with YAMLWorkflowParser. That parser resolves agent-level tools: declarations and preserves top-level approve:. Workflow.start() then installs those YAML-approved tools into the approval context. As a result, an untrusted recipe can omit execute_command from TEMPLATE.yaml requires.tools, declare it in workflow.yaml agents.*.tools, and add top-level approve: [execute_command]. The caller did not set allow_dangerous_tools=True, but the recipe policy allows the recipe and the workflow approval path self-approves the critical shell tool. This issue has been patched in version 4.6.61.
CVECVSSEPSSKEVRExploitTitleMod.
CVE-2026-539429.9 CRI—
———Conda Organization Infrastructure is a centralized location for infrastructure codes. Prior to commit 9c0f238, a GitHub Actions Pwn Request vulnerability appears to remain present in conda/conda. The affected workflow is .github/workflows/update.yml. It is triggered by issue_comment events containing @conda-bot render or @conda-bot recreate. For the render path, the workflow obtains the pull request head repository and branch from the PR API, checks out that PR-controlled repository/ref using secrets.SYNC_TOKEN, and then runs conda/actions/template-files@7f6830b1428a9bd47f0b068892c77eae95207037. The external template-files action reads its default configuration and templates from the checked-out working directory. In the confirmed finding, that action uses an unsandboxed Jinja2 environment to render attacker-controlled template content. This creates a Pwn Request chain in which an untrusted PR/comment path can reach template rendering in a privileged token context. This issue has been patched via commit 9c0f238.5hCVE-2026-399027.2 HIG—
———Cacti is an open source performance and fault management framework. Prior to version 1.2.31, the prior blocklist ([;&| `$\\n\r]) omitted single-quote, double-quote, redirect operators, subshell parens and braces. An authenticated admin could save an input_string` containing single-quote, redirect operator, or subshell parens, and pass the check. This issue has been patched in version 1.2.31.5hCVE-2026-115478——
———obsidian-tc is a governed, agent-ready Model Context Protocol server for Obsidian vaults. From 1.0.1 until 1.31.4, `create_entity`, `link_entities`, `rename_entity`, `unlink_entities` and `delete_entity` regenerate memory notes from SQLite at the computed `memory/<type>/<name>.md` path without reliably preserving note ownership or the consistency of a failed lifecycle operation. During shared materialization, a hand-authored note or an orphan left by a different entity whose name sanitized to the same path could have its entire body overwritten on the next write naming that type and name, with non-owned frontmatter preserved. `rename_entity` also seeded the destination path with a direct atomic write without validating ownership and could overwrite a foreign note there. Separately, several tools mutated SQLite before fallible filesystem work: `create_entity` and `link_entities` inserted the row or relation before materialization; `rename_entity` updated the entity, including a retired status, before updating its note; `unlink_entities` removed a relation before rematerializing the source note; and `delete_entity`'s cascade removed the target and its relations before rematerializing neighbouring notes. A later filesystem failure could therefore leave a row or relation changed and the memory graph inconsistent with the notes on disk even when the operation failed. An authorized caller with the required memory-tool permissions could, without intending to, destroy the body of a vault note outside the entity being written; `delete_entity` additionally requires its deletion permission and destructive confirmation, which this memory flaw does not bypass. Snapshots-before-write through `restore_note` mitigate the overwrite only where snapshots are enabled and the note had previously been written through obsidian-tc. This issue is fixed in version 1.31.4.5hCVE-2026-115477——
———obsidian-tc is a governed, agent-ready Model Context Protocol server for Obsidian vaults. From 1.13.0 until 1.31.4, on a Streamable HTTP deployment with `auth.jwtSecret` configured and a 2026-07-28-era client declaring `elicitation.form` support, a dispatch-gated tool call offered an `inputRequired` confirmation could still complete after the human declined or cancelled, or the elicitation response was omitted or malformed, provided the client echoed the validly bound `requestState` on resubmission. The server trusted an authentic, call-bound state as proof of approval even though that state was minted before the `elicitation/create` leg was asked and proved only that the resubmission belonged to the same call. A client behaving incorrectly or under an attacker's control could therefore execute dispatch-gated destructive vault operations without the human's approval, while still requiring the caller's ordinary tool permissions. Only tools marked `destructive: true` or requiring a scope with a human-in-the-loop floor were reachable through this bypass. The sixteen handler-side conditionally gated tools, including `write_note` overwrite and `move_note`, did not read elicitation state on the modern wire and let neither approval nor decline through that path. Affected deployments using stdio or leaving `auth.jwtSecret` unset instead used the CLI `elicit_token` fallback and did not offer this `requestState` round trip. The fix does not prevent a malicious client from reporting a fabricated approval, because the server cannot verify that the client displayed the form or that a human saw it. This issue is fixed in version 1.31.4.5hCVE-2026-1146236.5 MED—
———baserCMS is a website development framework. In versions 5.0.0 before 5.3.2 and 5.4.0 before 5.4.2, an authenticated user with only theme-file view permission can traverse outside the theme directory through the optional BcThemeFile plugin's Admin web screen. ThemeFilesController::parseArgs() constructs a path containing literal ../ segments, and the view and edit actions pass it to ThemeFilesService::get() without containment validation. Authorization checks the unresolved request path, after which ThemeFile::_getContents() reads the resolved file and exposes its contents in the form textarea. A client must preserve the dot segments rather than normalize them away. Readable files are limited to .php, .css, .js, and .ctp; this can disclose application source and config/install.php database credentials in MySQL or PostgreSQL deployments, while SQLite configurations have an empty database password. Reading config/app_local.php may expose the Security.salt fallback, which is effective only when not overridden by environment configuration. The flaw does not disclose .env, .pem, or SQLite .db files, does not itself permit file writes or code execution, and does not affect the separate API getFullpath() route. The plugin is disabled by default and unauthenticated requests are not sufficient. This issue has been patched in versions 5.3.2 and 5.4.2.4hCVE-2026-114355——
———OS command injection in the root-privileged 'hardware_monitor' helper in DJI CrystalSky ZS600C firmware v0.3.9.7 through v0.6.6.9 on Android allows a local unprivileged process to execute arbitrary commands as root by writing a crafted line to the Android log (logcat), which the helper reads and passes to a shell eval.4h