New: the Agent Incident Ledger: disclosed incidents of AI gone rogueExplore the ledger

Drive AI adoption, stay in control of the risks.

Agents are changing how every part of the business works, and they act with real access to code, data and production systems. Security has a double mandate: drive agent adoption and keep the risk under control.

To do both, security needs clear answers to four questions:

What agents are running?

Each agent arrives with its own entourage of MCP servers, skills, plugins and extensions, and every one of them adds to what the agent can do.

What can they touch?

Each agent holds an employee’s access to code, data and production, and loads a supply chain of MCP servers, skills and plugins nobody reviewed.

What are they doing at runtime?

Agents do real work. One prompt becomes dozens of shell commands, file edits and API calls, and security needs to see each action and know it stays within acceptable bounds.

How can we respond?

Some agents are turned by a malicious tool; others run away, reaching their goal by a route nobody wanted. Either way, security needs policies that catch the violation and act on it before it becomes an incident.

A hefty cost for organisations.

Security is buried in alerts and tool requests

Every new agent, MCP server and skill becomes a ticket, on top of the alerts. Approval queues run to weeks, and the backlog only grows.

Rogue agents are causing real damage.

Agent incidents are on the rise.

Deleted databases, leaked credentials, sandbox escapes. We track every publicly reported one in the Agent Incident Ledger.

Explore the Agent Incident Ledger
JUL 2025 · The RegisterReplit agent deletes a production database during a code freeze
JUL 2025 · BleepingComputerAmazon Q extension ships with an injected wiper prompt
AUG 2025 · The Hacker NewsCompromised Nx packages use AI CLIs to hunt for secrets
FEB 2026 · Tom’s HardwareClaude Code runs terraform destroy on production
FEB 2026 · TechCrunchOpenClaw bulk-deletes a researcher’s inbox and ignores stop
APR 2026 · The RegisterCursor agent deletes a production volume and its backups
APR 2026 · SecurityWeekMeta’s AI support assistant resets Instagram accounts for attackers
SEP 2026 · Help Net SecurityHijacked coding assistant spreads a worm across 100 repos
SEP 2026 · DecryptMeta’s Muse reads a columnist’s iMessages after he said no

The consequence: security is a bottleneck.

Caught in the middle, security has two bad options.

BAD OPTION 1

Say no, and block

Security doesn’t know what to block and what to allow, so aggressive policies stop developers and every other team using AI. A huge productivity loss, at a time when transformation is most needed.

  • Queues, reviews and bans for every new tool
  • People route around you, and shadow AI grows
  • The business falls behind in the AI race
BAD OPTION 2

Say yes, and shoulder an unknown risk

Agents spread across the stack with real access. Nobody can say what they touch, and the first sign that something went wrong is the incident itself.

  • Agents run with real access, unwatched
  • The first sign of trouble is the incident
  • Security owns a risk it can’t see
Manifold

Option 3: Drive AI transformation,
and stay in control.

Manifold uses the full context of every agent session.We instrument the harness where each agent runs, and connect the security stack you already have, from identity and EDR to MDM.

The widest agent coverage on the market.

160+agents and harnesses

  • Claude Code
  • Gemini CLI
  • Rovo Dev
  • Pi
  • Replit
  • Snowflake Cortex Code
  • Selenium Agent
  • Docker Model Runner
  • n8n
  • OpenAI Codex
  • Qwen Code
  • GitHub Copilot CLI
  • Grok Bot
  • Muse
  • Trae
  • VS Code
  • Ollama
  • Cursor
  • Windsurf
  • Kiro
  • Perplexity
  • Junie
  • Warp
  • VSCodium
  • OpenShell

And many more.

Every policy you need, from a nudge to a full stop.

Not every risky action deserves a block. Set the response by group, agent, tool or action: watch quietly, guide the agent, fix the call, ask the user, or stop it cold. Policies are written as code and enforced inside the harness, in under 200ms.

  1. WATCH

    Alert

    Record it for security. The user carries on uninterrupted.

  2. NUDGE

    Advise

    Let it run, and steer the agent onto a safer path.

  3. FIX

    Correct

    Rewrite the action into a safe one, then let it run.

  4. CONFIRM

    Ask

    The user approves or denies it. No answer, no action.

  5. STOP

    Block call

    Stop this one action, with the reason sent back to the agent.

  6. END

    Block session

    End the agent session. Every step stays on record.

LIGHTER TOUCHFIRMER CONTROL

Threat intelligence from Manifest, built in.

Manifest is our AI supply chain intelligence. Everything your agents load is checked against it, alongside a live threat feed curated by Manifold Research.

698,571assets vetted, and counting: MCP servers, skills, plugins and editor extensions

Search Manifest

See Manifold in action.

A short tour of the platform, from the first agent discovered to the first policy enforced.