GHSA-fc26-m9pf-v56q: PraisonAI LinearBot: webhook auth bypass enables agent exec

# PraisonAI LinearBot processes unsigned webhooks when `LINEAR_WEBHOOK_SECRET` is missing ## Summary PraisonAI's LinearBot starts a public webhook listener on `0.0.0.0` and treats `LINEAR_WEBHOOK_SECRET` as optional. When the secret is absent, startup only logs a warning and `_handle_webhook()` skips `Linear-Signature` verification entirely. An unauthenticated network caller who can reach the webhook endpoint can submit a forged `Linear-Event: AgentSession` request. The forged request is parsed, scheduled for background processing, dispatched to `_handle_agent_session()`, and passed into `BotSessionManager.chat()`. The bot then attempts to post the agent response back to Linear under the configured bot token. The local PoV is offline and deterministic. It does not contact Linear. It calls the webhook handler directly, monkey-patches the outbound Linear comment path, and proves both sides of the boundary: - no secret configured: unsigned forged webhook returns `200`, invokes the agent session path once, and attempts one Linear comment; - secret configured: missing and bad signatures both return `401` and do not invoke the agent; - secret configured with valid HMAC: request returns `200` and invokes the agent, proving the control path still works. ## Affected Product - Repository: `MervinPraison/PraisonAI` - Package: `praisonai` - Components: - `src/praisonai/praisonai/bots/linear.py` - `src/praisonai/praisonai/cli/features/bots_cli.py` Validated affected: - live `main` / latest observed release `v4.6.58`: `1ad58ca02975ff1398efeda694ea2ab78f20cf3e` - previous local current checkout: `2f9677abb2ea68eab864ee8b6a828fd0141612e1` - `v4.6.57` - `v4.6.56` - `v4.5.50` Sampled tags where the LinearBot component was not present: - `v4.5.49` - `v4.5.51` - `v4.6.9` - `v4.6.10` Suggested affected range: LinearBot-bearing releases with the fail-open signature behavior, at least `4.5.50` and `>= 4.6.56, <= 4.6.58`. The component appears non-contiguously in sampled tags, so maintainers should confirm the exact packaged version history before publishing a final range. ## Root Cause `LinearBot.__init__()` accepts an empty signing secret and falls back to an empty environment value: ```python self._signing_secret = signing_secret or os.environ.get("LINEAR_WEBHOOK_SECRET", "") ``` `start()` treats the missing secret as a warning instead of refusing to expose the webhook listener: ```python if not self._signing_secret: logger.warning("LINEAR_WEBHOOK_SECRET not set - webhook signatures will not be verified") self._site = web.TCPSite(self._runner, "0.0.0.0", self._webhook_port) ``` `_handle_webhook()` only verifies the request if the secret is truthy: ```python if self._signing_secret: signature = request.headers.get("Linear-Signature", "") if not self._verify_signature(raw_body, signature): return web.Response(status=401, text="Invalid signature") ``` With no secret configured, the code continues to JSON parsing, accepts a caller supplied `webhookTimestamp`, reads the caller supplied `Linear-Event` header, and schedules processing: ```python event_type = request.headers.get("Linear-Event", "") task = asyncio.create_task(self._process_webhook(event_type, body)) return web.Response(status=200, text="OK") ``` For `AgentSession`, the forged body is routed to the agent: ```python if event_type == "AgentSession": await self._handle_agent_session(body) ... response = await self._session_mgr.chat(self._agent, user_id, message.content) await self._send_comment(...) ``` The CLI has the same fail-open posture: `start_linear()` loads `LINEAR_WEBHOOK_SECRET`, prints a warning when it is missing, then reports a public `http://0.0.0.0:<port>/webhook` endpoint with verification disabled. ## Why This Is Not Intended Behavior PraisonAI's Linear Bot documentation tells operators to set `LINEAR_WEBHOOK_SECRET`, pass it to `praisonai bot linear`, copy the Linear webhook signing secret, and use it for HMAC-SHA256 verification. The same page says missing secrets disable signature verification, while its best-practices section says webhook secrets ensure authenticity. Linear's webhook documentation says receivers should ensure requests were sent by Linear by verifying the `Linear-Signature` HMAC over the raw body, then checking that `webhookTimestamp` is recent. The timestamp check alone is not an authentication boundary because an attacker can supply a current timestamp in a forged body. The implementation itself also confirms the intended boundary: when a secret is configured, missing and bad signatures are rejected before agent dispatch. The bug is the missing-secret fail-open mode on a public webhook server, not the signature algorithm. ## Local PoV Run against the latest observed release checkout: ```bash python3 submission-bundle/praisonai-prai-cand-013-linear-webhook-signature-fail-open/poc/pov_prai_cand_013_linear_webhook_signature_fail_open.py --repo artifacts/repos/praisonai-v4.6.58 ``` Expected output includes: ```json { "candidate": "PRAI-CAND-013", "ok": true, "cases": { "no_secret_unsigned_forged_webhook": { "http_status": 200, "signing_secret_configured": false, "session_calls": [ { "user_id": "linear-system", "content": "Issue: Forged Linear AgentSession event\n\nPRAI-CAND-013 local forged webhook payload" } ], "sent_comments": [ { "issue_id": "issue-prai-cand-013", "comment": "agent response", "session_id": "prai-cand-013-session" } ] }, "secret_missing_signature_control": { "http_status": 401, "session_calls": [] }, "secret_bad_signature_control": { "http_status": 401, "session_calls": [] }, "secret_valid_signature_control": { "http_status": 200, "session_calls": [ { "user_id": "linear-system" } ] } } } ``` Stored evidence: - `evidence/pov-v4.6.58.json` - `evidence/pov-live-main-v4.6.58.json` - `evidence/pov-current-head.json` - `evidence/version-sweep.tsv` ## Impact If a PraisonAI operator starts LinearBot with a Linear token but omits `LINEAR_WEBHOOK_SECRET`, any network caller that can reach the webhook endpoint can spoof Linear webhook events and invoke the configured agent through the Linear integration. For the `AgentSession` event path, this lets the attacker supply issue title and description content that becomes the agent input. Depending on the configured agent and tools, this can cause unauthorized LLM/tool execution, consume paid model quota, create or update Linear comments under the bot identity, and drive the bot into workflows intended only for authenticated Linear events. This report does not claim arbitrary code execution by default. The concrete boundary crossed is unauthenticated remote agent invocation through a forged Linear webhook. ## Suggested Fix Fail closed for public webhook listeners: 1. Refuse to start LinearBot when `LINEAR_WEBHOOK_SECRET` is missing, unless an explicit development-only option such as `--insecure-skip-webhook-signature-verification` is provided. 2. In `_handle_webhook()`, reject requests when no signing secret is configured instead of silently skipping verification. 3. Preserve raw-body HMAC verification and constant-time comparison for the configured-secret path. 4. Treat timestamp freshness as replay protection after signature validation, not as a replacement for authentication. 5. Prefer loopback binding by default, or require an explicit host flag for public binding. 6. Add regression tests: - no signing secret rejects startup or rejects webhook requests; - missing signature with a configured secret returns `401`; - invalid signature with a configured secret returns `401`; - valid HMAC with a configured secret returns success; - stale timestamp after valid HMAC returns `401`; - the CLI does not advertise a public unauthenticated webhook by default.

An attacker discovers a PraisonAI LinearBot instance via port scan or leaked CI/CD configuration and confirms the webhook is unauthenticated by sending a probe POST that returns 200. The attacker crafts a malicious POST to /webhook with the Linear-Event: AgentSession header and a JSON body containing a forged issue title and description embedding prompt injection instructions — for example, directing the agent to enumerate configuration files, exfiltrate environment variables, or invoke a code-execution tool. The handler, lacking a signing secret, skips all signature validation, parses the body, and dispatches it to BotSessionManager.chat() with the attacker-controlled content as the agent's user message. The agent executes the injected prompt, posts its response to a Linear issue under the bot token, and — if configured with file-system or external-API tools — may chain the injection into broader system-level actions, all without any authentication step.