Advisory

GHSA-h2w2-v7j6-xqm4

Also known as: CVE-2026-57137

Summary

npm PraisonAI AgentLoop onToolCall approval runs after tool execution

Details

Summary

The published npm package praisonai exports createAgentLoop(), whose onToolCall callback is documented and exampled as an approval hook. The implementation calls PraisonAI's generateText() wrapper with the caller's executable tools first, receives toolResults, and only then calls onToolCall().

Because AI SDK generateText() executes tools with an execute function as part of the generation call, onToolCall can deny a tool only after the sensitive side effect has already happened. PraisonAI then returns finishReason: "tool_rejected", which is a false security signal: the rejected tool already ran.

The PoV is deterministic and local-only. It uses mock AI SDK modules, no live model call, no API key, and no network target. The tool increments an in-memory counter rather than touching the filesystem or executing commands.

Technical Details

In src/praisonai-ts/src/ai/agent-loop.ts, the public config says:

/** On tool call callback (for approval) */
onToolCall?: (toolCall: ToolCallInfo) => Promise<boolean>;

The inline approval example also asks a user for approval and returns the decision:

onToolCall: async (toolCall) => {
  const approved = await askUserForApproval(toolCall);
  return approved;
}

However, AgentLoop.step() calls generateText() with the executable tools before invoking onToolCall:

const result = await generateText({
  model: this.config.model,
  messages: this.messages as any,
  tools: this.config.tools,
  maxSteps: 1,
});

It then materializes toolResults:

toolResults: result.toolResults.map(tr => ({
  toolCallId: tr.toolCallId,
  toolName: tr.toolName,
  result: tr.result,
})),

Only afterward does the approval callback run:

if (this.config.onToolCall) {
  for (const toolCall of step.toolCalls) {
    const approved = await this.config.onToolCall(toolCall);
    if (!approved) {
      this.complete = true;
      step.finishReason = 'tool_rejected';
      break;
    }
  }
}

src/praisonai-ts/src/ai/generate-text.ts forwards the caller's tools directly to AI SDK:

const result = await sdk.generateText({
  model,
  ...
  tools: options.tools,
  maxSteps: options.maxSteps,
  ...
});

AI SDK documents that generateText() "generates text and calls tools", and that tools with an execute function run automatically unless approval is handled before execution with needsApproval.

The published npm:[email protected] dist files preserve the same order:

  • dist/ai/agent-loop.js lines 150-157 call generateText() with executable tools.
  • lines 162-171 materialize toolResults.
  • lines 183-195 call onToolCall() and set tool_rejected afterward.

Why This Is Not Intended Behavior

This is not a trust-model-only issue. PraisonAI explicitly labels onToolCall as an approval callback and shows an approval example. A user who returns false from that callback expects the tool not to run.

It also conflicts with the AI SDK execution model PraisonAI wraps:

  • AI SDK generateText() executes tools that include an execute function.
  • AI SDK approval is a pre-execution boundary (needsApproval), not a post-execution notification.
  • AI SDK loop control documentation treats "a tool call needs approval" as a condition that stops or pauses the loop before executing the tool.

PraisonAI's current behavior instead creates a post-execution audit hook while naming and documenting it as approval.

PoV

Run from a local reproduction checkout:

node poc/pov_poc.js 1.7.1

Expected output includes:

{
  "praisonaiVersion": "1.7.1",
  "createAgentLoopExported": true,
  "eventOrder": ["tool-executed", "approval-denied"],
  "sideEffects": 1,
  "finishReason": "tool_rejected",
  "toolCallCount": 1,
  "toolResultCount": 1,
  "rejectedAfterExecution": true,
  "vulnerable": true,
  "patchedControl": {
    "order": ["approval-denied"],
    "sideEffects": 0,
    "toolCallCount": 1,
    "toolResultCount": 0,
    "blocksBeforeExecution": true
  }
}

The PoV installs npm:[email protected] into a temporary project and supplies mock ai and @ai-sdk/openai modules. The mocked generateText() returns one tool-call intent and executes a supplied execute handler if present. This keeps the proof deterministic and isolates PraisonAI's ordering bug.

The vulnerable run uses createAgentLoop() with:

  • a dangerousWrite tool whose execute() handler increments an in-memory side-effect counter and records tool-executed;
  • an onToolCall approval callback that always returns false and records approval-denied.

The observed order is:

tool-executed > approval-denied

That proves denial happens after execution. The toolResults array contains the tool's result even though PraisonAI reports finishReason: "tool_rejected".

The patched-control comparison strips executable handlers before the model step, requests approval on the tool-call intent, and only executes if approval succeeds. With the same denial decision, the control output is:

approval-denied
sideEffects = 0
toolResultCount = 0

PoC

The PoV section above contains the local reproduction command, input, and decisive output.

Impact

Any application using npm PraisonAI createAgentLoop() with onToolCall as a human-in-the-loop or policy approval boundary can execute denied tools.

If the application exposes the agent loop to lower-trust prompts or users and registers powerful tools, an attacker can cause the model to call a tool that the approval callback denies. The denial occurs too late. Depending on the registered tool, impact can include file modification, command execution, external API calls, data mutation, credential use, or other side effects with the privileges of the PraisonAI process.

The report does not claim that npm PraisonAI exposes this as a default network service. It is a library-level approval-boundary bypass in the exported TypeScript agent-loop API.

Severity

Suggested severity: High.

Rationale:

  • AV: common deployment pattern is an application exposing agent prompts over a network.
  • AC: attacker only needs to induce a tool call.
  • PR: conservative base score assumes the attacker can submit prompts to the application.
  • UI: no additional operator action is needed for the tool to execute before denial; even a denial callback is too late.
  • S: impact is in the PraisonAI-hosting application process.
  • C/I/A: depends on registered tools; shell/file/API tools can affect confidentiality, integrity, and availability.

If maintainers score only local scripts that process untrusted repositories or prompts, AV:L may be reasonable. If they score public unauthenticated prompt endpoints built on this API, PR:N may be reasonable.

Suggested Fix

Do not pass executable tool handlers into generateText() before approval.

One safe shape:

  1. Convert configured tools into intent-only tool definitions without execute.
  2. Call generateText() to obtain the model's tool-call intent.
  3. Invoke onToolCall(toolCall) before any side effect.
  4. Execute the selected tool only if approval returns true.
  5. Append approved tool results to the conversation and continue the loop.

Severity

High Basis: The source's own band: HIGH

Dates

Published
2026-06-18 14:26 UTC
Modified
2026-07-20 21:56 UTC

Affected packages and fixes

praisonai

npm · All systems

Affected

  • introduced 1.4.0, fixed 1.7.2 (semver)

Fixed versions

1.7.2

Fix

Fix commands are the export's own generic commands. Ojuju never runs them. Check them against your setup first.

Check your own computer: guard packages check

View this record on OSV.dev

Source

Source: GitHub Advisory Database (reviewed advisories) · CC BY 4.0

Sources & licences