New47 coding-agent incidents and findings, sourced
Endpoint security for AI agents

Every AI agent. Every risky action, stopped before it runs.

  • See every AI agent and MCP server your teams use, and who runs each one.
  • Stop destructive commands and reads of credential files before they happen, in the operating system, below the agent.
  • Keep a record of what each agent did, session by session.
Rolled out through your MDM·Runs next to your EDR·Streams to your SIEM
Live decisionson the device, below the agent
  1. Claude Codep.nair · paymentsread ~/.aws/credentials
    Block
  2. Claude Codep.nair · paymentsGitHub token in a prompt to a model API
    Redact
  3. Cursors.rao · platformrm -rf ~/work/payments-svc
    Block
  4. Codexa.menon · datarun migration on prod
    Ask r.iyer
  5. Claude Codep.nair · paymentsopen pull request #482
    Allow
  6. Gemini CLIk.das · growthload mcp: unknown-tool
    Block
Allowed24Blocked6Redacted7Asked a person4Illustrative example
Securing agents at
Mysa TagMango AI Fiesta Fleapo
Member of the NVIDIA Inception program
The challenge

Every laptop is becoming an agent host.

AI agents now act on company devices with real access. Most security teams can't see them, can't limit them and can't trace what they did.

82%of enterprises have unknown AI agents in their environment
53%have had AI agents exceed their intended permissions
12.5%can consistently trace an agent from instruction to outcome

Sources: Cloud Security Alliance agent security surveys, 2026. Salt Security, State of Agentic AI and API Security, 2H 2026.

Case study

In 19 days on one company's laptops, GIRD watched AI agents send out 4 API keys.

Observation mode, so nothing was blocked, and nothing looked like an incident. 8 of 14 installed coding agents never ran a session, and week one held 12.7 million agent actions.

Read the case study
4keys sent out inside agent requests, two GitHub and two Shopify
8 of 14installed coding agents with no recorded session
12.7Magent actions recorded in week one
216sends of one unidentified token to a model API in 16 hours
The platform

One control plane for your whole agent estate

GIRD runs where your agents run, sees what they do from the operating system up, and decides every action against one policy. Everything lands in one record.

How GIRD works

From discovery to evidence

Find every agent, decide every action, keep data on the device, and hand auditors the record.

Discover

See every agent, including the ones nobody approved.

An inventory of every AI agent, MCP server and model in use, each tied to a person and a team.

  • Shadow agents flagged the day they appear
  • Dormant agents found before an attacker uses them
  • Agents, tools and models in one registry
Agent maplast 7 days
1,284 agents37 unapproved212 MCP servers100% with an ownerIllustrative
Decide

Check every action against the task it was given.

Reading a test file fits a task to fix a test. Reading cloud credentials does not, so it never runs. Policy runs in the operating system, so an agent with its hooks turned off is governed all the same.

  • Allow, block, redact or ask a person in Slack or Teams
  • Least privilege by team and by task
  • No round trip to a cloud for a decision
Claude Code sessionp.nair on payments
TaskFix the flaky refund test in payments
  1. read payments/refund_test.goIn task
  2. go test ./payments/...In task
  3. edit payments/refund.goIn task
  4. read ~/.aws/credentialsBlocked
  5. curl paste.example.netBlocked
  6. git push origin fix/refund-flakeIn task
fits the task

Steps 4 and 5 have nothing to do with fixing a test, so they never ran.

Protect

Keep secrets and regulated data on the laptop.

Keys, national ID numbers such as Aadhaar, PAN and US Social Security numbers, card numbers and confidential documents are caught on the device before a prompt leaves it. Documents are matched by meaning, not file name.

  • Secrets and tokens from every major provider
  • Redacted on the device, so prompts never leave to be checked
  • Your confidential documents, matched by meaning
Outbound requestClaude Code, support tools

Refund order 88213 for priya.s@example.com, card 4111 1111 1111 1111, PAN ABCDE1234F. Use key sk_live_9f2c41aa and attach Q3 board pack.pdf.

Sent with 5 items redacted on the device

Illustrative example
Prove

Hand auditors a record, not a story.

Every decision is sealed to the one before it, tied to a person, an agent and a policy, and streamed to Splunk, Microsoft Sentinel, Datadog or Google SecOps. Controls map to SOC 2, ISO 27001, ISO 42001, NIST AI RMF and regional rules.

  • Answer which agent did this, and who ran it, in seconds
  • Evidence packs ready for auditors
  • Retention set to your rules, kept in your region

Chain verified. 3 of 3 seals match.

14:02:11 · Claude Code · p.nairBlockedread ~/.aws/credentialsseal 9c41..e2 · prev 7a03..51
14:02:14 · Claude Code · p.nairRedactedGitHub token in a promptseal 3be8..07 · prev 9c41..e2
14:09:40 · Codex · a.menonApproved by r.iyermigration on prod paymentsseal f0d2..9a · prev 3be8..07
Streams to Splunk · Sentinel · Datadog · Google SecOpsIllustrative example
The rules

The rules already ask you to watch AI and limit what it can touch.

Lawmakers, regulators and standards bodies expect monitoring, least-privilege access and human oversight for AI systems. GIRD gives you the control on every device and the evidence to show it.

EU AI Act

Law
Deployers “shall monitor the operation” of high-risk AI systems, assign human oversight to people with the authority to act, and keep the system's logs for “at least six months”.
Article 26. Applies to high-risk systems from 2 December 2027.
With GIRDEvery agent action is monitored on the device, high-impact actions go to a named approver, and a sealed record is kept for as long as you set.

CERT-In

Advisory · India
AI agents reached live systems of Indian organisations through exposed credentials and over-permissive cloud roles, and “none of the affected organizations detected the access”. CERT-In recommends a least-privilege identity for each agent run, outbound access denied by default and a tamper-resistant log of agent activity.
Advisory CIAD-2026-0049, 15 October 2026, rated Critical.
With GIRDAgents can't read credential files, outbound access is limited to the destinations you allow, and every file, command and connection is recorded.

OWASP Top 10 for LLM Applications

Standard
For excessive agency, OWASP says to “limit the permissions that LLM extensions are granted to other systems to the minimum necessary”, and “to require a human to approve high-impact actions before they are taken”.
LLM06:2025, Excessive Agency.
With GIRDEach agent and MCP tool is scoped by team and by task, and a person is asked in Slack or Teams before high-impact actions run.

NIST SP 800-53

Control
AC-6 calls for least privilege, allowing only the accesses that users, “or processes acting on behalf of users”, need for their assigned tasks.
AC-6, Least Privilege.
With GIRDAn AI agent is a process acting on behalf of a user. GIRD checks each action against the task the agent was given and blocks the rest.

GIRD supports these requirements. It does not by itself make an organisation compliant. Check with your counsel which obligations apply to you.

Why GIRD

Inventory tells you what is installed. GIRD decides what it does.

Below the agent

Enforced in the operating system, not inside the agent. Hooks off or no hooks at all, the action is still seen and decided.

Where the keys live

GIRD runs on each laptop, right where agents open files and keys. It sees a local read of ~/.aws/credentials and stops it before it happens.

Aware of the task

Every step is scored against what the developer asked for, so a step that has nothing to do with the task stands out at once.

Built for regulators

Built for the EU AI Act's monitoring and log duties, CERT-In's advisory, and OWASP and NIST least-privilege guidance. Evidence stays in the region you choose.

Where GIRD fits next to what you already run

EDR protects the machine. Gateways see network traffic. Inventory tools list what is installed. None of them decide what an agent is allowed to do, step by step, on the device.

CapabilityEDRAI gateway or proxyEndpoint inventory toolsGIRD
Finds every agent, MCP server and modelPartlyNoYesYes
Decides each agent action before it runsNoPartlyNoYes
Knows the task the agent was givenNoNoNoYes
Governs agents that bypass the networkPartlyNoPartlyYes
Redacts secrets and PII before they leave the deviceNoPartlyNoYes
Tamper-evident record of every agent decisionPartlyPartlyNoYes

Based on what each category typically does. Individual products vary.

Rollout

GIRD on every laptop, rolled out by IT.

Your keys and code live on each device, so GIRD runs there too, beside your AI agents. IT pushes it through your MDM, so engineers have nothing to install or configure.

  1. First

    IT pushes it through your MDM

    Jamf, Intune, Kandji or Workspace ONE install GIRD on every laptop. Mac builds are signed and notarized by Apple.

  2. Then

    Watch before you block

    Observation mode maps every agent and learns how each one normally works, blocking nothing.

  3. When you're ready

    Turn on policy

    By team and by task. If GIRD itself stops, your teams keep working.

Device managementJamf, Microsoft Intune, Kandji, Workspace ONE
IdentityOkta, Microsoft Entra ID, Google Workspace
SIEM and SOARSplunk, Microsoft Sentinel, Datadog, Google SecOps
Runs next to your EDRCrowdStrike Falcon, SentinelOne, Microsoft Defender
Approvals and alertsSlack, Microsoft Teams, Jira, ServiceNow
Trust

Built for regulated teams, by people who have been audited.

Checked on the device

Actions are decided where they run, without sending files or prompts to a cloud first.

Your data, your region

Choose where records live and how long they stay.

Tamper-evident record

Every decision is sealed into a chain you can verify.

Never stalls engineering

If GIRD itself stops, your teams keep working.

Signed builds

Rolled out through your MDM. Mac builds are notarized by Apple.

Security reviews

We complete your CAIQ, SIG or own questionnaire.

Disclosure

We acknowledge reports within 2 business days and fix high-severity issues within 30.

Trust Center

What we collect and never collect, where it lives, and who processes it, in one place.

Controls map toSOC 2ISO/IEC 27001ISO/IEC 42001NIST AI RMFOWASP Agentic Top 10MITRE ATLASCERT-InDPDP Act
Company

The team

Akash Shaw

Akash Shaw

Co-founder · CEO

Seven years building the systems regulators audit. At CRED he owned a lending platform doing 100,000+ loans a day under RBI supervision, and let AI agents write its production code for nine months under guardrails he wrote.

Anagha Joshi

Anagha Joshi

Co-founder · CTO

Fourteen years of engineering leadership in regulated fintech, ecommerce and healthcare. At CRED she took Motor Insurance, Garage and BNPL from zero to one.

FAQ

Questions security teams ask first.

Will this slow my engineers down?

GIRD runs locally on each laptop and decides there, with no round trip to a cloud. IT rolls it out through your MDM, so engineers have nothing to install, and if GIRD itself stops, your teams keep working.

Do we need to replace our EDR?

No. GIRD runs alongside CrowdStrike Falcon, SentinelOne and Microsoft Defender. Your EDR protects the machine. GIRD decides what each AI agent on it is allowed to touch.

What if an agent has no hooks, or a developer turns them off?

Policies run in the operating system, below the agent. An agent that reports nothing is governed the same way as one that reports everything.

Do prompts or code leave our devices?

Detection runs on the device, so a prompt never leaves it to be checked. You choose the region where records live and how long they are kept.

Which agents are covered?

Coding agents such as Claude Code, Cursor, Codex, GitHub Copilot and Gemini CLI, plus desktop assistants, browser agents, MCP servers and the agents your teams build.

Which systems does it run on?

Macs today, rolled out through your MDM. Windows and Linux are next.

Can we try it before we buy?

Yes. Pilots start in observation mode, so nothing is blocked until you choose. To look at one Mac first, the Argus community edition runs on its own.

How is GIRD priced?

Per machine, per month, billed monthly or yearly. Book a risk review and we'll send a quote for your fleet.

Book a risk review

See what your agents did this week.

30 minutes with a founder. You leave with a map of your exposure and a first policy, whether or not you buy. Pilots start in observation mode, so nothing is blocked until you choose.

  1. 10 min
    Map your agent estateWhich agents, MCP servers and models your teams run, and who owns each one.
  2. 10 min
    Walk through a real sessionHow GIRD decides each action, and what it would have stopped in the 19-day case study.
  3. 10 min
    Draft your first policyWhat to block, what to redact and what needs a person, mapped to the rules you report on.
Akash ShawCo-founder · CEO
Anagha JoshiCo-founder · CTO