AI AppSec Assistant

​

Overview

​

Key Capabilities

​

Autonomous Security Operations

• Triage findings — assess true/false positive status, evaluate severity, recommend remediation priority
• Investigate vulnerabilities — clone repos, review code paths, assess reachability
• Create patches — generate fixes, create PRs through the platform’s patch pipeline
• Threat modeling — analyze application architecture and identify attack surfaces
• SCA reachability — determine whether dependency vulnerabilities are actually exploitable in your code

​

Agent Teams

​

Self-Refinement

• Memory — stores and recalls decisions, team preferences, repo-specific knowledge, and investigation results across runs. Before each task, it searches memory for prior decisions on similar issues to stay consistent. After completing work, it saves key outcomes so future runs don’t re-investigate the same issues.
• Natural language rules — when it identifies a recurring pattern (e.g., repeated false positives from a specific code pattern), it can create scanner rules so the platform catches or suppresses similar findings automatically in future scans. This means the assistant’s triage decisions permanently improve the scanner’s accuracy.
• Recurring task context — for organization-wide scheduled tasks, the assistant automatically receives the results of its last 3 runs. It uses this to track trends, build on prior work, and avoid starting from scratch each time.

​

Event-Driven Automation

• A scan completes → the assistant triages new findings and posts a summary
• A PR is blocked → the assistant reviews the blocking findings and suggests fixes
• A new CVE is discovered → the assistant assesses reachability across your repos
• A scan fails → the assistant investigates the error and notifies the team

​

Persistent Environment

• Recurring task context — organization-wide schedules can reuse the previous three runs’ results to track trends. Scoped schedules reuse a conversation only while its resolved repository selection stays identical; each run gets a separate file workspace.
• Cross-run memory — organization-wide tasks can store and retrieve organizational knowledge. Scoped schedules do not access shared organization memory.

​

Controllability

• Global instructions: org-wide directives prepended to every task. Set policies like “always post in Slack before modifying records” or “escalate critical findings to @security-oncall.” Administrators edit them in Configure → Settings or through the administrative API. MCP tools cannot change them.
• Event triggers — configure which platform events (scan complete, new vulnerabilities, etc.) automatically invoke the assistant, with customizable prompts per event type.
• Cron scheduling — the assistant can create and manage its own recurring schedules (e.g., “triage new findings every Monday at 9am”).
• Permissions — three granular permissions control who can view, manage, or run the assistant.

​

Dashboard

• Overview: the Jobs running, Jobs completed today, Active schedules, and Active triggers counts and whether Slack is connected, a Running now list of jobs in progress with live duration counters, and Recent actions, a feed with AI-generated summaries of the concrete actions taken (for example, findings triaged, PRs opened, scans triggered).
• Activity: a searchable list of all conversations with AI-generated titles that summarize each conversation’s purpose. You can filter by source (Chats, Issues, PRs, Scheduled, Events, Slack) and click any row to open the full conversation thread with real-time streaming.
• Chat: start new conversations or continue existing ones. The conversation history lists all conversations grouped by day (Today, Yesterday, then by date) with search and source filtering (All, Chat, Issue, PR, Scheduled, Event, Slack). On narrower screens the conversation history collapses behind a Conversation history toggle button. Conversations from scheduled tasks and event triggers are read-only, while Slack-linked conversations direct you to reply in the original Slack thread. The assistant automatically receives the full conversation history when responding to follow-up messages, so it maintains context across multiple exchanges in the same thread. A Stop button cancels all currently pending or running jobs in the conversation at once, which halts a multi-step task mid-flight. Stopping a conversation with no active jobs does nothing. If the Assistant is switched off for your organization, the chat input is disabled and displays a message indicating this; an admin can re-enable it under Assistant → Configure → Settings.
• Configure: the Assistant’s configuration sections, described below.

​

Configure

• Schedules: create and manage recurring tasks with cron expressions, templates, and repo/tag scoping.
• Triggers: configure which platform events invoke the assistant, with per-trigger prompts and repo/tag scoping.
• Playbooks: browse and manage reusable playbook templates that define multi-step workflows the assistant can execute. Playbooks codify common security operations (for example, triage runbooks or incident response procedures) so they run the same way across your team.
• Skills: manage the reusable know-how the assistant reaches for while it works. Built-in skills are enabled automatically and can be disabled or duplicated into an editable copy, and New skill adds a custom one. Changes apply to the next run.
• Memory: view, search, edit, and delete the assistant’s stored organizational knowledge.
• Settings: turn the assistant on or off, set the Slack channel for updates, set global instructions, and restart the assistant.

​

Quick chat from any page

• With no conversation open, the drawer offers starting prompts such as Triage the newest critical issues.
• Conversation history lists your conversations, and New conversation starts a fresh one.
• Open in Assistant opens the current conversation on the Assistant page.
• Closing the drawer keeps your draft and the conversation, so reopening it picks up where you left off.

​

Getting Started

​

Enable the Assistant

1. Navigate to Assistant in the ZeroPath dashboard.
2. Open Configure, then Settings.
3. Turn on the switch under Assistant status.

​

Run Your First Task

1. Click the Ask button in the top right, or go to the Chat tab.
2. Type a message in the input at the bottom of the page to start a new conversation, for example:
   • “List all critical severity findings from the last scan and triage them”
   • “Review the latest PR scan results and flag any true positives”
   • “What are the top 5 most common vulnerability types across our repos?”
3. Watch the results stream in real-time. You can send follow-up messages in the same conversation thread.

​

Set Up Global Instructions

1. Open Assistant → Configure → Settings and find Global instructions.
2. If the field is empty, click Load example instructions for a starting template.
3. Write your instructions. For example:

## Escalation & Approval

- Before modifying or deleting any issue, post a summary in Slack
  and wait for confirmation from a team member before proceeding.
- For critical or high severity findings, escalate to @security-oncall
  in #security-alerts with finding details and your recommended action.

## Memory

- At the start of every task, search memory for prior decisions about
  similar findings. Use existing decisions as precedent.
- After completing any triage or investigation, store the key decision
  and reasoning in memory for future reference.

## Pull Requests

- Never auto-merge pull requests. Create the PR, tag the repo owner,
  and post the PR link in Slack for review.

4. Click Save instructions.

​

Event Triggers

​

Configuring Triggers

1. Open Configure → Triggers.
2. Click “Add trigger” on the event type you want, or toggle on an existing trigger:

3. Click “Edit prompt” to customize what the assistant does for each event.
4. Optionally scope each trigger to specific repositories or tags so it only fires for a subset of your codebase.

​

Inbound Webhooks

​

Creating an Inbound Webhook

1. Open Configure → Triggers.
2. Under the Webhooks section, click “Add trigger”.
3. Give the trigger a name and write a prompt template. Use {webhookBody} as a placeholder for the incoming request payload.
4. Click “Save”. ZeroPath generates a unique webhook URL with a secret token.

POST https://zeropath.com/api/agent/webhook/{triggerId}?secret={webhookSecret}

​

Example

A new bug bounty report was submitted. Analyze the report for validity,
check if the described vulnerability exists in our codebase, and post
your assessment to Slack.

Report contents:
{webhookBody}

{
  "title": "Open redirect in /auth/callback",
  "severity": "medium",
  "description": "The redirect_uri parameter is not validated..."
}

​

Scoping

​

Trusted Mode

• Trusted — organization-wide runs can use shell access, file read/write, and configured MCP tools. Use this only when the prompt content is controlled by your team.
• Untrusted (default) — the assistant runs with no shell access and blocked access to sensitive configuration files. Use this for triggers that process external input, such as inbound webhooks carrying payloads from outside your organization.

​

When to Use Untrusted Mode

• Inbound webhooks receiving payloads from external systems
• Triggers processing bug bounty reports, vendor advisories, or customer-submitted data
• Any trigger where {webhookBody} or other placeholders carry externally-authored text

​

What Changes in Untrusted Mode

​

Prompt Variables

​

Scheduled Tasks

​

Creating Schedules from the UI

1. Open Configure → Schedules and click “New Schedule”.
2. Choose a template to start from (Daily Triage, Weekly Security Digest, or SCA Monitoring), or build one from scratch.
3. Enter a name, select a cron expression (or use one of the built-in presets like “Weekdays at 9am UTC” or “Every hour”), and write the prompt that tells the assistant what to do.
4. Optionally scope the schedule to specific repositories or tags so the assistant only operates on a subset of your codebase.
5. Click Save. You can enable, disable, edit, or trigger any schedule manually from the list.

​

Creating Schedules via Chat

• “Create a weekly triage of all new findings every Monday at 9am”
• “Check for critical SCA vulnerabilities daily at midnight”
• “Run a security posture summary every Friday at 4pm”

​

Failure Notifications

​

Recurring Task Context

• Trend tracking — “Last week there were 12 critical findings, this week there are 8. Improvement in repos X and Y.”
• Incremental work — the assistant can pick up where it left off rather than starting from scratch.
• Audit trail — every run’s output is preserved for review.

​

Slack Integration

​

Setup

1. Ensure your organization has a Slack integration connected.
2. Open Assistant → Configure → Settings, enter the channel ID under Slack updates, and click Save.
3. The assistant will respond to @-mentions in channels and direct messages.

​

Talking to the Assistant

@ZeroPath Can you triage the 5 new critical findings from today’s scan?

What’s the current security posture across our repos?

​

Supported Actions

​

Jira & Confluence Integration

• Create remediation tickets linked to ZeroPath findings
• Look up existing tickets for context when triaging
• Search Confluence for architecture documentation to inform threat modeling and reachability analysis

​

Linear Integration

• Create remediation tickets directly from triage results
• Query existing tickets for remediation context

​

PR Workflow for Linear Tickets

• Includes the Linear ticket key (for example, ZP-123) in the branch name, PR title, and PR body so Linear automatically links the PR to the ticket.
• Moves the Linear ticket to Todo after posting the PR link, so the ticket owner can review and verify the fix. The ticket is not moved to a completed or closed state — it stays in Todo until you verify the fix and close it manually.

​

Custom MCP Servers

​

Memory

• Triage decisions and reasoning
• Team preferences and security policies
• Repository-specific knowledge (architecture, common patterns, known issues)
• Remediation plans and outcomes

​

Viewing and Managing Memory

• Search memories by keyword to find specific entries.
• Edit any memory entry inline to correct or refine what the assistant has stored.
• Delete individual entries that are no longer relevant.
• Clear all memories to reset the assistant’s organizational knowledge entirely. This action cannot be undone.

​

Best Practices

• The assistant automatically searches memory before each task and stores outcomes after. You don’t need to tell it to remember things explicitly.
• If you want to correct a past decision, tell the assistant directly — it will update its memory.
• Use global instructions to guide memory behavior (e.g., “always check memory for prior decisions on similar CVEs before triaging”).

​

Permissions

​

Repository-Level Access Control

​

Audit Logging

​

API Access

export ZEROPATH_TOKEN_ID='your-token-id'
export ZEROPATH_TOKEN_SECRET='your-token-secret'
export ZEROPATH_ORG_ID='your-organization-id'
export ZEROPATH_BASE_URL='https://zeropath.com'

Content-Type: application/json
X-ZeroPath-API-Token-Id: your-token-id
X-ZeroPath-API-Token-Secret: your-token-secret

1. Read settings with POST /api/v2/agent/getConfig:
   { "organizationId": "your-organization-id" }
   
   The response contains config, with effective settings and nullable savedConfig metadata. Reading does not create configuration and never returns webhook secrets.
2. Enable the Assistant with POST /api/v2/agent/updateConfig:
   { "organizationId": "your-organization-id", "enabled": true }
   
   Optional fields are slackChannelId, globalInstructions, maxConcurrentRuns, and chatRetentionDays. Omitted fields stay unchanged. Authorize Slack in Settings → Integrations before setting its channel. Configuration is available through administrative API tokens and the dashboard, not MCP tools.
3. Create an untrusted, tag-scoped schedule with POST /api/v2/agent/createCronSchedule:
   {
     "organizationId": "your-organization-id",
     "name": "weekly-tag-review",
     "cronExpression": "0 9 * * 1",
     "prompt": "Review new security findings in the selected repositories.",
     "tagIds": ["11111111-1111-4111-8111-111111111111"],
     "trusted": false
   }
   
   This runs Mondays at 09:00 UTC. Save the returned schedule.id. trusted is the requested setting; effectiveTrusted is false for repository- and tag-scoped schedules.
4. Run it immediately with POST /api/v2/agent/triggerCronSchedule:
   { "organizationId": "your-organization-id", "scheduleId": "returned-schedule-id" }
   
   Save the returned jobId. An immediate run keeps the recurring cadence unchanged and is available for paused schedules. A disabled Assistant fails with HTTP 412.
5. Poll POST /api/v2/agent/getJob with that ID:
   { "organizationId": "your-organization-id", "jobId": "returned-job-id" }
   
   resultSummary contains a brief outcome. Complete reports from scoped scheduled runs are saved as ASSISTANT messages before temporary run files are removed. job.status becomes COMPLETED, FAILED, or CANCELLED. The response includes resultSummary, customer-visible messages, and the run’s effective trusted value. Pending or running jobs can have a null summary. Results respect current organization, authorship, and repository permissions.