Back to skill

Security audit

inbound-qualification

Security checks for vulnerabilities and agentic risk

Overview

The skill’s inbound-lead workflow is coherent, but it needs review because public form text is posted into an internal Slack channel as markdown without sanitization.

Review before installing or deploying. Keep the explicit deploy approval, origin allowlist, and privacy-page requirement, and fix the Slack message path by escaping or rendering visitor-provided values as plain text with server-side length limits.

Vulnerability Patterns
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
Findings (1)

T09 · Insecure Skill Coding Practices

Warning
Location
infra/tools/inbound-form.ts:150
Finding

Public Form Content Is Injected into Internal Slack Markdown

Content
View full analysis

Vulnerability Details

File Location: infra/tools/inbound-form.ts, lines 150–156
Vulnerability Type: Untrusted Slack Markdown Injection
Risk Level: Medium

ts
uses.slack.postMessage({
  channelId: slackChannelId,
  format: "markdown",
  disableUnfurling: true,
  body: `*New inbound${qualified ? ", qualified" : ""}* from ${input.first_name} ${input.last_name} at ${company.company_name ? company.company_name : domain} (${domain})\n${company.employee_count} employees, ${company.hq_country}\n${input.message}\n\nPage: ${input.page_url}`,
});

Technical Analysis

The public workflow accepts visitor-controlled first_name, last_name, and message strings through the schema at infra/tools/inbound-form.ts:59–71. These values have no relevant character restrictions or length limits and are interpolated directly into a Slack message configured with format: "markdown".

Consequently, Slack interprets rather than merely displays supported markup supplied by an external visitor. An attacker can submit formatting, deceptive links, and Slack mention constructs that alter the appearance or notification behavior of the internal message. The trusted prefix does not separate attacker-controlled content into a plain-text field.

Setting disableUnfurling: true only controls link previews; it does not escape Markdown, neutralize links, or suppress mention parsing. The configured origin allowlist, honeypot, minimum-fill timer, and rate limit reduce automated abuse but do not sanitize a submission accepted from the legitimate website.

This is an ordinary exploitable coding flaw. The reviewed evidence does not indicate that the project author intentionally introduced the injection behavior.

Attack Path

  1. An external attacker opens the legitimate public contact page, satisfying the configured origin restriction.
  2. The attacker waits longer than minFillMillis and submits a syntactically valid non-free ema ...[truncated 1192 chars]
Remediation
View remediation

Remediation Suggestions

  • Send all visitor-controlled fields through Slack plain-text blocks or another API field that does not interpret Markdown.
  • If Markdown is required for trusted labels, escape Slack control characters in every untrusted value before interpolation and neutralize mention and link constructs.
  • Keep trusted formatting separate from visitor-provided values rather than constructing one Markdown string.
  • Add server-side maximum lengths for names, messages, and URL-derived fields to limit notification abuse.
  • Add tests using mention syntax, formatted links, markup delimiters, and oversized values, verifying that Slack renders them as inert text.
  • Retain the existing origin restriction, honeypot, minimum-fill timer, and rate limit as defense-in-depth; they should not replace output encoding.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (13)

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 104)May include surrounding context.

md
5. **Plan, then stop.** `node --import tsx evals/contract.mjs` from this skill's folder, then

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 130)May include surrounding context.

md
5. **Plan, then stop.** `node --import tsx evals/contract.mjs` from this skill's folder, then

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 177)May include surrounding context.

md
5. **Plan, then stop.** `node --import tsx evals/contract.mjs` from this skill's folder, then

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 116)May include surrounding context.

md
| ICP rules (`infra/tools/inbound-form.ts`) | derived | the ICP in `context/`: headcount band and ISO country codes, confirmed with one live enrichment | Decide

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 152)May include surrounding context.

md
| ICP rules (`infra/tools/inbound-form.ts`) | derived | the ICP in `context/`: headcount band and ISO country codes, confirmed with one live enrichment | Decide

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 144)May include surrounding context.

md
rch` | The form's rules and the Slack post are enough | Delete `infra/plays/research-qualified-leads.ts`, `infra/agents/lead-researcher.ts` and the Anthropic co

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 164)May include surrounding context.

md
rch` | The form's rules and the Slack post are enough | Delete `infra/plays/research-qualified-leads.ts`, `infra/agents/lead-researcher.ts` and the Anthropic co

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The trigger list includes very broad phrases such as 'handle inbound leads' and 'add a demo form to our website', which can cause an agent to invoke this skill in situations beyond the narrowly intended website-form workflow. That increases the chance of misrouting user requests into a flow that deploys public forms, connectors, Slack notifications, and data writes without the user explicitly asking for this specific architecture.

Content

No source excerpt is available for this finding.

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · SKILL.md (reported line 114)May include surrounding context.

md
**Derive before you ask.** An input with a lookup is looked up, not asked.

| Input                                     | Kind    | How it is answered                                                                              | Why it matters                                                                                            |
| ----------------------------------------- | ------- | ----------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- |
| ICP rules (`infra/tools/inbound-form.ts`) | derived | the ICP in `context/`: headcount band and ISO country codes, confirmed with one live enrichment | Decides who sees the booking link. Rules nobody can read back are the reason inbound stops being trusted. |
| site origin (`publicForm.allowedOrigins`) | derived | the website app's `site.json` `canonicalUrl`, without the trailing slash                        | Every other origin is refused with 403. A missing origin is a form that never submits.                    |

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · SKILL.md (reported line 137)May include surrounding context.

md
The code is a worked example. These reshapes are expected, and the agent offers them rather than
waiting to be asked. Every one costs something.

| Variation               | When it is right                                          | How                                                                                                                                          | What it costs                                                                              |
| ----------------------- | --------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
| `crm-backed`            | Contacts belong in HubSpot, Salesforce or Attio           | Replace the two `model.upsert` calls with the CRM connector's upsert (matched on domain and email), same fields ([data](references/data.md)) | The worked example no longer deploys on a bare workspace; CRM properties must exist first. |
| `agent-qualification`   | The ICP does not reduce to headcount and country          | Replace the rules with an agent that reads the ICP from `context/` and returns a verdict and a reason                                        | Each submission pays for a model call, and the answer is less predictable.                 |

Whitespace Padding

Medium
Category
Prompt Injection
Confidence
70% confidence
Finding

Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.

Content

Scanner excerpt · SKILL.md (reported line 143)May include surrounding context.

md
| `agent-qualification`   | The ICP does not reduce to headcount and country          | Replace the rules with an agent that reads the ICP from `context/` and returns a verdict and a reason                                        | Each submission pays for a model call, and the answer is less predictable.                 |
| `owner-routing`         | More than one rep takes inbound                           | Assign `owner_id` on the contact by territory or round robin before the Slack post, and mention the owner                                    | An owner table to keep current, and a fallback when nobody matches.                        |
| `turnstile`             | Spam gets through the honeypot, time-trap and rate limit  | `publicForm.spam.captchaProvider: "turnstile"`, the site key, `captchaSecret: env("TURNSTILE_SECRET")`, the widget on the page               | A third-party script on the page, with its own privacy disclosure.                         |
| `accept-personal-email` | A form that is not about the company (newsletter, events) | Drop the free-mail refusal, skip enrichment for those domains                                                                                | No company to qualify or route; those contacts arrive without an account.                  |
| `no-deep-research`      | The form's rules and the Slack post are enough            | Delete `infra/plays/research-qualified-leads.ts`, `infra/agents/lead-researcher.ts` and the Anthropic connector                              | No tier or brief on the contact; the team researches by hand.                              |
| `research-every-submitter` | Not-qualified leads are worth a look too               | Drop the `inbound_status` condition from the research play's filter                                                                         | An agent call per submission, including the ones the rules already turned away.            |

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The document describes collecting, storing, syncing, and enriching personal data from form submissions, including names, email addresses, page URL, UTM parameters, message content, and marketing consent, and potentially sending that data into shared models, Slack, and CRM systems. Without an explicit user-facing privacy notice, consent basis, retention guidance, and data-sharing disclosure, this creates a real privacy and compliance risk because operators may deploy the flow in a way that processes personal data beyond what the submitter reasonably expects.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
87% confidence
Finding

The submit handler collects first name, last name, email, message, and consent, then passes them to the Cargo form SDK and submits them asynchronously. Although comments describe the behavior for developers, the user-facing form does not disclose that this data will be transmitted to a backend service and may trigger downstream processing.

Content

No source excerpt is available for this finding.