GHSA-v847-hxxw-3pxg CVE-2026-56838 PYSEC-2026-3520

PraisonAI recipe.run_stream skips dangerous-tool policy enforcement

High PyPI

Severity: 7.8 / 10; High · CVSS 3.1
Exploit chance (EPSS): 0.15%; Chance of exploitation in the next 30 days. Higher than 3.84% of scored vulnerabilities.
Known exploited: No; Not in CISA's catalog
Weakness: CWE-693 Protection Mechanism Failure; CWE-78 Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection'); CWE-863 Incorrect Authorization
Fix: 4.6.59; Upgrade praisonai to this version or later

Description

PraisonAI `recipe.run_stream()` skips dangerous-tool policy enforcement

Summary

PraisonAI recipe execution blocks default-denied dangerous tools unless the caller explicitly passes allow_dangerous_tools=True. The normal recipe.run() path enforces this with _check_tool_policy(). The streaming path, recipe.run_stream(), loads the same recipe, checks dependencies, and then calls _execute_recipe() without running the dangerous-tool policy check.

As a result, a recipe that honestly declares execute_command in TEMPLATE.yaml requires.tools is denied by recipe.run(), but reaches the execution engine through recipe.run_stream() with allow_dangerous_tools=False.

The local PoV uses a harmless printf canary, explicitly unsets PRAISONAI_AUTO_APPROVE, and avoids network access.

Affected Product

Repository: MervinPraison/PraisonAI
Package: praisonai
Components:
- src/praisonai/praisonai/recipe/core.py
- src/praisonai/praisonai/recipe/serve.py
- src/praisonai/praisonai/cli/features/recipe.py
- src/praisonai-agents/praisonaiagents/workflows/yaml_parser.py
- src/praisonai-agents/praisonaiagents/workflows/workflows.py

Validated affected:

current main 2f9677abb2ea68eab864ee8b6a828fd0141612e1 (v4.6.57-4-g2f9677ab)
v4.6.57
v4.6.56
v4.6.10
v4.6.9
v4.5.128
v4.5.120
v4.5.96
v4.5.87

Suggested affected range: >= 4.5.87, <= 4.6.57.

PyPI lists PraisonAI 4.6.57 as the latest release on 2026-06-13.

Earlier tested tags through v4.5.85 failed in this source checkout before the tested workflow path due an unrelated praisonaiagents.output.models import error. They are not claimed fixed or unaffected.

Root Cause

recipe.run() enforces the dangerous-tool gate:

if not options.get("allow_dangerous_tools", False):
    policy_error = _check_tool_policy(recipe_config)
    if policy_error:
        return RecipeResult(..., status=RecipeStatus.POLICY_DENIED, ...)

recipe.run_stream() has a sibling execution path. It loads the recipe and checks dependencies, but then goes directly to execution:

recipe_config = _load_recipe(name, offline=options.get("offline", False))
...
output = _execute_recipe(recipe_config, merged_config, session_id, options)

There is no equivalent _check_tool_policy() call in run_stream() before execution or before the dry-run shortcut.

The CLI exposes this path via praisonai recipe run <recipe> --stream, and the recipe HTTP server exposes it as POST /v1/recipes/stream.

Why This Is Not Intended Behavior

The normal recipe path clearly treats declared dangerous tools as denied by default. A control recipe with TEMPLATE.yaml requires.tools: [execute_command] returns:

Tool 'execute_command' is denied by default. Use allow_dangerous_tools=True to override.

That operator-facing override should not depend on whether the caller requests streaming output. PraisonAI's own docs describe approval as requiring a human or configured channel before risky tools run, describe security environment variables as opt-in access for dangerous operations with secure defaults, and describe policy controls as blocking dangerous operations.

This is distinct from the prior report PRAI-CAND-011:

PRAI-CAND-011 covers workflow tool declarations that are omitted from TEMPLATE.yaml requires.tools.
This report covers a sibling entrypoint that skips the policy check even when TEMPLATE.yaml correctly declares the dangerous tool.

It is also distinct from the published Recipe-server authentication fail-open advisory. That advisory covers missing authentication secrets. This report assumes the attacker has whatever access is already needed to invoke recipe streaming and focuses on the missing dangerous-tool policy guard.

Local PoV

Run:

python3 poc/pov_prai_cand_012_stream_policy_bypass.py

Expected output includes:

{
  "ok": true,
  "policy_error": "Tool 'execute_command' is denied by default. Use allow_dangerous_tools=True to override.",
  "control_recipe_status": "policy_denied",
  "execution_reached": [
    {
      "recipe": "declared-dangerous-stream",
      "declared_required_tools": ["execute_command"],
      "allow_dangerous_tools": false
    }
  ],
  "workflow_approve_tools": ["execute_command"],
  "runner_tool_names": ["execute_command"],
  "command_stdout": "PRAI-CAND-012-CANARY",
  "operator_env_auto_approve": null
}

The PoV creates a temporary recipe that declares execute_command in TEMPLATE.yaml requires.tools.

Control:

recipe.run(..., options={"force": True}) returns policy_denied.

Bypass:

recipe.run_stream(..., options={"force": True}) emits the executing event and reaches _execute_recipe() while allow_dangerous_tools remains false.
The same recipe workflow resolves execute_command and preserves approve: [execute_command].
With the workflow approval context installed, the resolved tool runs the harmless local command printf PRAI-CAND-012-CANARY.

The PoV monkey-patches _execute_recipe() only to prove that run_stream() crosses the policy boundary without invoking an LLM. The command canary is executed directly through the same resolved workflow tool and approval context to keep the proof deterministic and local-only.

Impact

If an operator runs an untrusted recipe through streaming mode, or exposes the recipe streaming API to users who can choose recipe names or URIs, the recipe can reach execution with default-denied tools even though the caller did not set allow_dangerous_tools=True.

If the workflow reaches the approved execute_command tool call, commands run with the privileges of the PraisonAI process. The exact trigger depends on the workflow and model/tool-call path, but the dangerous-tool policy boundary is already bypassed before execution.

The HTTP recipe sidecar is documented as a localhost REST API with SSE streaming and optional API-key/JWT authentication. This report does not claim default unauthenticated network RCE. In authenticated or exposed sidecar deployments where lower-trust users can invoke /v1/recipes/stream, the same policy gap can become a remote recipe-execution issue.

Suggested Fix

Centralize recipe preflight enforcement so every execution mode uses the same guard:

Run _check_tool_policy(recipe_config) in run_stream() unless options["allow_dangerous_tools"] is true.
Perform that check before both dry-run and real execution, matching recipe.run().
Prefer a shared helper for dependency checks, dangerous-tool policy checks, and dry-run handling so future entrypoints cannot drift.
Add regression tests:
- declared dangerous tool is denied by recipe.run();
- the same declared dangerous tool is denied by recipe.run_stream();
- allow_dangerous_tools=True preserves the intended opt-in behavior;
- /v1/recipes/stream maps a policy denial to a non-success SSE event or equivalent HTTP failure.

Affected versions

PyPI

Package	Vulnerable	Fixed in
praisonai	`>= 4.5.87, < 4.6.59`	`4.6.59`

How to fix

Move to the fixed version or later with your package manager, update the lockfile, and run your tests. If the package only arrives through another dependency, upgrade that dependency first. If none of its releases pulls a fixed version yet, force the version with your package manager's override.

praisonai PyPI → `4.6.59`

pip

pip install --upgrade "praisonai>=4.6.59"

poetry

poetry add "praisonai>=4.6.59"

uv

uv add "praisonai>=4.6.59"

Severity breakdown

CVSS 3.1 7.8

CVSS:3.1/AV:L/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H

Attack vector: Local
Attack complexity: Low
Privileges required: None
User interaction: Required
Scope: Unchanged
Confidentiality: High
Integrity: High
Availability: High

Weakness

CWE-693: Protection Mechanism Failure

The product does not use or incorrectly uses a protection mechanism that provides sufficient defense against directed attacks against the product.

CWE-78: Improper Neutralization of Special Elements used in an OS Command ('OS Command Injection')

The product constructs all or part of an OS command using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the intended OS command when it is sent to a downstream component.

How to prevent it

Architecture and Design: If at all possible, use library calls rather than external processes to recreate the desired functionality.
Architecture and Design, Operation: Run the code in a "jail" or similar sandbox environment that enforces strict boundaries between the process and the operating system. This may effectively restrict which files can be accessed in a particular directory or which commands can be executed by the software. OS-level examples include the Unix chroot jail, AppArmor, and SELinux. In general, managed code may provide some protection. For example, java.io.FilePermission in the Java SecurityManager allows the software to specify restrictions on file operations. This may not be a feasible solution, and it only limits the impact to the operating system; the rest of the application may still be subject to compromise. Be careful to avoid CWE-243 and other weaknesses related to jails.
Architecture and Design: For any data that will be used to generate a command to be executed, keep as much of that data out of external control as possible. For example, in web applications, this may require storing the data locally in the session's state instead of sending it out to the client in a hidden form field.
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.
Architecture and Design: Use a vetted library or framework that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. For example, consider using the ESAPI Encoding control [REF-45] or a similar tool, library, or framework. These will help the programmer encode outputs in a manner less prone to error.
Implementation: While it is risky to use dynamically-generated query strings, code, or commands that mix control and data together, sometimes it may be unavoidable. Properly quote arguments and escape any special characters within those arguments. The most conservative approach is to escape or filter all characters that do not pass an extremely strict allowlist (such as everything that is not alphanumeric or white space). If some special characters are still needed, such as white space, wrap each argument in quotes after the escaping/filtering step. Be careful of argument injection (CWE-88).
Implementation: If the program to be executed allows arguments to be specified within an input file or from standard input, then consider using that mode to pass arguments instead of the command line.
Architecture and Design: If available, use structured mechanisms that automatically enforce the separation between data and code. These mechanisms may be able to provide the relevant quoting, encoding, and validation automatically, instead of relying on the developer to provide this capability at every point where output is generated. Some languages offer multiple functions that can be used to invoke commands. Where possible, identify any function that invokes a command shell using a single string, and replace it with a function that requires individual arguments. These functions typically perform appropriate quoting and filtering of arguments. For example, in C, the system() function accepts a string that contains the entire command to be executed, whereas execl(), execve(), and others require an array of strings, one for each argument. In Windows, CreateProcess() only accepts one command at a time. In Perl, if system() is provided with an array of arguments, then it will quote each of the arguments.
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, denylists can be useful for detecting potential attacks or determining which inputs are so malformed that they should be rejected outright. When constructing OS command strings, use stringent allowlists that limit the character set based on the expected value of the parameter in the request. This will indirectly limit the scope of an attack, but this technique is less important than proper output encoding and escaping. Note that proper output encoding, escaping, and quoting is the most effective solution for preventing OS command injection, although input validation may provide some defense-in-depth. This is because it effectively limits what will appear in output. Input validation will not always prevent OS command injection, especially if you are required to support free-form text fields that could contain arbitrary characters. For example, when invoking a mail program, you might need to allow the subject field to contain otherwise-dangerous inputs like ";" and ">" characters, which would ne
Architecture and Design: When the set of acceptable objects, such as filenames or URLs, is limited or known, create a mapping from a set of fixed input values (such as numeric IDs) to the actual filenames or URLs, and reject all other inputs.
Operation: Run the code in an environment that performs automatic taint propagation and prevents any command execution that uses tainted variables, such as Perl's "-T" switch. This will force the program to perform validation steps that remove the taint, although you must be careful to correctly validate your inputs so that you do not accidentally mark dangerous inputs as untainted (see CWE-183 and CWE-184).
Operation: Run the code in an environment that performs automatic taint propagation and prevents any command execution that uses tainted variables, such as Perl's "-T" switch. This will force the program to perform validation steps that remove the taint, although you must be careful to correctly validate your inputs so that you do not accidentally mark dangerous inputs as untainted (see CWE-183 and CWE-184).
Implementation: Ensure that error messages only contain minimal details that are useful to the intended audience and no one else. The messages need to strike the balance between being too cryptic (which can confuse users) or being too detailed (which may reveal more than intended). The messages should not reveal the methods that were used to determine the error. Attackers can use detailed information to refine or optimize their original attack, thereby increasing their chances of success. If errors must be captured in some detail, record them in log messages, but consider what could occur if the log messages can be viewed by attackers. Highly sensitive information such as passwords should never be saved to log files. Avoid inconsistent messaging that might accidentally tip off an attacker about internal state, such as whether a user account exists or not. In the context of OS Command Injection, error information passed back to the user might reveal whether an OS command is being executed and possibly which command is being used.
Operation: Use runtime policy enforcement to create an allowlist of allowable commands, then prevent use of any command that does not appear in the allowlist. Technologies such as AppArmor are available to do this.
Operation: Use an application firewall that can detect attacks against this weakness. It can be beneficial in cases in which the code cannot be fixed (because it is controlled by a third party), as an emergency prevention measure while more comprehensive software assurance measures are applied, or to provide defense in depth [REF-1481].
Architecture and Design, Operation: Run your code using the lowest privileges that are required to accomplish the necessary tasks [REF-76]. If possible, create isolated accounts with limited privileges that are only used for a single task. That way, a successful attack will not immediately give the attacker access to the rest of the software or its environment. For example, database applications rarely need to run as the database administrator, especially in day-to-day operations.
Operation, Implementation: When using PHP, configure the application so that it does not use register_globals. During implementation, develop the application so that it does not rely on this feature, but be wary of implementing a register_globals emulation that is subject to weaknesses such as CWE-95, CWE-621, and similar issues.

CWE-863: Incorrect Authorization

The product performs an authorization check when an actor attempts to access a resource or perform an action, but it does not correctly perform the check.

How to prevent it

Architecture and Design: Divide the product into anonymous, normal, privileged, and administrative areas. Reduce the attack surface by carefully mapping roles with data and functionality. Use role-based access control (RBAC) [REF-229] to enforce the roles at the appropriate boundaries. Note that this approach may not protect against horizontal authorization, i.e., it will not protect a user from attacking others with the same role.
Architecture and Design: Ensure that access control checks are performed related to the business logic. These checks may be different than the access control checks that are applied to more generic resources such as files, connections, processes, memory, and database records. For example, a database may restrict access for medical records to a specific database user, but each record might only be intended to be accessible to the patient and the patient's doctor [REF-7].
Architecture and Design: Use a vetted library or framework that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid. For example, consider using authorization frameworks such as the JAAS Authorization Framework [REF-233] and the OWASP ESAPI Access Control feature [REF-45].
Architecture and Design: For web applications, make sure that the access control mechanism is enforced correctly at the server side on every page. Users should not be able to access any unauthorized functionality or information by simply requesting direct access to that page. One way to do this is to ensure that all pages containing sensitive information are not cached, and that all such pages restrict access to requests that are accompanied by an active and authenticated session token associated with a user who has the required permissions to access that page.
System Configuration, Installation: Use the access control capabilities of your operating system and server environment and define your access control lists accordingly. Use a "default deny" policy when defining these ACLs.

References

Advisories

Packages

Timeline

Jun 18, 2026 Published
Oct 29, 2026 Last updated by the source
Nov 1, 2026 Latest EPSS score

Advisory	Severity	Why it's related
GHSA-26mh-57q7-jfvr PraisonAI vulnerable to Remote Code Execution via Broken AST Sandbox	High	Also affects praisonai
GHSA-6g6r-q6gw-w8fg PraisonAI has a Browser Server WebSocket origin validation bypass via unanchored regex (patch bypass of CVE-2026-40289 / GHSA-8x8f-54wf-vv92)	Critical	Also affects praisonai
GHSA-pvph-5j39-v8qc PraisonAI: Origin-validation bypass (startswith prefix match) enables unauthenticated cross-site request forgery against the PraisonAI MCP HTTP server	High	Also affects praisonai
GHSA-gfq8-hmph-9gjv PraisonAI: Authentication fail-open in Recipe server allows unauthenticated access when API key or JWT auth is configured without a secret	High	Also affects praisonai
GHSA-2jgc-f764-c5r2 PraisonAI: [Auth Bypass] PraisonAI async Jobs API (`/api/v1/runs`) has no authentication — unauthenticated job execution, result theft, cancel and delete	High	Also affects praisonai
GHSA-pvxx-r596-f5qj PraisonAI: `--api-key` flag on `praisonai serve` is not properly enforced	High	Also affects praisonai
GHSA-hmfx-4v44-9qw9 PraisonAI vulnerable to Server-Side Request Forgery via DNS rebinding bypass in webhook_url validation	Medium	Also affects praisonai
GHSA-rg5q-pp8p-f7jm PraisonAI: Webhook SSRF via DNS fail-open in `JobSubmitRequest.validate_webhook_url()` — bypass of CVE-2026-40114	High	Also affects praisonai

Is this in your code?

MergeHarbor checks every repository's dependencies against this database, ranks what matters by exploitability and reachability, routes it to the owning team and opens the fix pull request.

Check your repos with MergeHarbor