Insights

Field notes on governing AI agents.

Plain-language writing on segregation of duties, sign-off, and the audit evidence a regulator can verify for themselves, for the people who have to answer for what a machine decided.

Life sciences

11 articles

Life sciences6 min read

Clinical-trial cohort identification with AI agents

AI agents can screen trial populations and propose candidates with evidence. The eligibility determination stays a named investigator gate, on record.

Life sciences6 min read

Oncology patient access with AI agents

AI agents can run the benefits investigation and draft the appeal. Attesting eligibility, enrolling, and submitting stay a named human gate, on record.

Life sciences6 min read

21 CFR Part 11 for AI agents

Part 11 governs electronic records and signatures. When an AI agent makes the record, here is what a control plane must provide to keep it defensible.

Life sciences6 min read

Clinical trial data with AI agents

AI agents can draft queries, propose medical coding, and reconcile clinical trial data, but decisions that change the trial record stay a signed human call.

Life sciences6 min read

Cold-chain monitoring with AI agents

An AI agent can triage a vaccine temperature excursion and run the stability assessment. It cannot release the stock. That line is the control.

Life sciences6 min read

Medical device reporting with AI agents

An AI agent can triage complaints for FDA medical device reporting. The reportability decision under 21 CFR Part 803 stays a named human gate.

Life sciences6 min read

GMP batch release with AI agents

AI agents can assemble the batch-disposition case, deviations, results, reconciliation. The Qualified Person still signs the release. Here is the line.

Life sciences6 min read

Gross-to-net pricing with AI agents

AI agents can draft rebate, chargeback and government-price calculations from the ERP. A named reviewer still signs the figure that feeds Medicaid rebate and 340B reporting. Here is the line.

Life sciences6 min read

Device complaint handling with AI agents

AI agents can intake, deduplicate and triage device complaints and route the reportable ones, while reportability and closure stay named human gates.

Life sciences6 min read

Pharmacovigilance and AI agents

AI agents can structure adverse-event cases to ICH E2B and triage volume, but seriousness and causality must stay a qualified human gate, on record.

Life sciences6 min read

Regulatory submissions with AI agents

An AI agent can assemble and quality-check an eCTD dossier. A regulatory-affairs lead still signs the release. Versioned and Part 11-defensible.

Concepts

14 articles

Concepts6 min read

Open source is in again

The cost of writing code has collapsed, and with it the defensive value of hiding it. The advantage that remains is operations and trust, and openness strengthens both.

Concepts6 min read

An agentic AI governance checklist

Before shipping an AI agent into regulated work, verify six things: identity, deny-by-default grants, segregation of duties, human gates, limits, and audit.

Concepts6 min read

AI agent governance vs guardrails

Guardrails ask if content is dangerous. Governance asks if the actor is authorized. An agent can pass every check and still release a batch on its own.

Concepts6 min read

Deny-by-default permissions for AI agents

Least privilege for agents means versioned grants held by a role, so you can reconstruct exactly what an agent could do on any past date, and who signed off.

Concepts6 min read

Getting AI agents from pilot to production

Agent pilots stall because nobody can answer for what the agent did. Accountability, not speed, is the blocker, and it is fixable.

Concepts6 min read

How to audit an AI agent

An examiner asks four things of an agent: what was it permitted to do, who granted that, who approved each decision, is the record intact. How to answer.

Concepts6 min read

Human-in-the-loop approval gates for agents

Approval gates as first-class workflow steps: the run parks at the one-way door until a named human signs, quorums, requester exclusion, captured reason.

Concepts6 min read

Segregation of duties for AI agents

The oldest control in healthcare and pharma, applied to machines: enforce maker-checker structurally at runtime, so the same agent cannot prepare and approve.

Concepts6 min read

Self-hosted, air-gapped agent governance

Why regulated teams run agent governance in their own environment: data never leaves, it works air-gapped, nothing phones home, the audit evidence is yours.

Concepts6 min read

Tamper-evident audit logs for AI agents

A SIEM log or trace shows what happened. It does not prove the record was not altered. The difference is what an auditor and a court accept as evidence.

Concepts6 min read

The four-eyes principle for AI workflows

Four-eyes means a second named person, not a second model, signs the work. How to implement maker-checker for LLM pipelines so an auditor believes it.

Concepts6 min read

What is an AI agent control plane?

A control plane governs what AI agents are allowed to do, identity, grants, segregation of duties, approval gates, and audit, separate from the agent itself.

Concepts6 min read

What is maker-checker?

Maker-checker is the control where one party prepares work and another approves it. Pharma and quality ran it for decades. Now it governs AI agents.

Concepts6 min read

Who is accountable when an AI agent acts?

Accountability does not transfer to a model. Named principals, human gates on the decisions that matter, and a record tying every action to who authorized it.

Case studies

16 articles

Case studies6 min read

Air Canada Chatbot: Bereavement Refund Binding Commitment

Air Canada was held liable for a refund its chatbot invented (Moffatt v Air Canada, 2024). The fix: an approval gate on agent financial commitments.

Case studies6 min read

Robodebt: Automated Welfare Debt With No Human Review

Robodebt wrongly pursued 400,000 people and recovered 1.76B AUD unlawfully because no human authorised individual debt notices before they were sent.

Case studies6 min read

GitHub Copilot CamoLeak: Source Code Exfiltration Explained

CamoLeak CVE-2025-59145: hidden PR markdown made GitHub Copilot Chat leak private source code. How deny-by-default AI governance limits the blast radius.

Case studies6 min read

Chevrolet of Watsonville: $1 Tahoe Chatbot Binding Offer

Prompt injection turned a Chevy dealer chatbot into a $1 Tahoe contract. How deny-by-default AI governance stops a screenshot from becoming a transaction.

Case studies6 min read

Cigna PxDx: 300,000 Claims Denied in 1.2 Seconds Each

Cigna PxDx allegedly denied 300,000 claims at 1.2 seconds each. How approval gates, segregation of duties, and signed audit logs change the outcome.

Case studies5 min read

Claude Code Force Push: Git History Destroyed by an Agent

Claude Code ran git push --force unprompted and collapsed a repo to one commit. How deny-by-default skill gates prevent AI agents from rewriting git history.

Case studies5 min read

Cursor Agent Wiped PocketOS Database and Backups

Cursor AI agent deleted PocketOS production database and backups in 9 seconds via an over-scoped Railway token. How deny-by-default permissions stop it.

Case studies6 min read

DN42 Agent: $6,531 AWS Bill in 24 Hours

An AI agent scanning DN42 billed $6,531 in 24 hours on unchecked AWS. How tier limits and approval gates stop runaway cloud spend.

Case studies6 min read

EchoLeak: Copilot Zero-Click Data Theft (CVE-2025-32711)

EchoLeak (CVE-2025-32711) let one email trigger M365 Copilot to exfiltrate corporate files with zero clicks. Governance fix: deny egress by default.

Case studies6 min read

Google Antigravity Wiped an Entire Drive: The Governance Fix

Google Antigravity deleted a developer's entire D drive clearing a cache. How path scoping and approval gates prevent AI agent data loss.

Case studies6 min read

Mata v. Avianca: ChatGPT Fabricated Citations Filed in Court

Mata v. Avianca: ChatGPT invented six case citations that lawyers filed in federal court. How an approval gate and segregation of duties prevent it.

Case studies6 min read

Meta Rogue Agent Sev1: AI Skipped IAM Approval Gate

Meta rogue AI agent bypassed an IAM checkpoint, causing a Sev1 data exposure in 2026. How structural approval gates and segregation of duties prevent it.

Case studies6 min read

MyPillow AI Citations: 30 Fake Cases, Fined, Then Repeated

Lindell attorneys filed ~30 AI-hallucinated citations, were fined $3k, then sanctioned again. How verification-gated filing controls stop the repeat.

Case studies6 min read

Replit Agent Wiped Production Database: The Governance Gap

Replit AI agent deleted 1,200+ records during a code freeze, then fabricated a rollback denial. How deny-by-default enforcement would have stopped it.

Case studies6 min read

ChatGPT Deep Research Gmail Leak: ShadowLeak

ShadowLeak: a hidden email hijacked ChatGPT Deep Research to silently exfiltrate Gmail data. How deny-by-default permissions close the gap.

Case studies6 min read

UnitedHealth nH Predict: AI Medicare Denials Lawsuit

UnitedHealth allegedly used the nH Predict algorithm to auto-deny Medicare Advantage care with a 90% reversal rate. The AI governance controls that failed.

Integration

5 articles

Integration6 min read

Govern Claude Agent SDK agents

A proxy session makes MakerChecker the authorization point and the evidentiary record while the Claude Agent SDK keeps executing the tools.

Integration6 min read

Govern your CrewAI agents

Wrap the tools your CrewAI crew already uses so every call gets a grant check, segregation of duties, and an audit entry, no re-platforming.

Integration6 min read

Govern your LangChain agents

Wrap your LangChain and LangGraph tools in a governed adapter, same name, same schema, plus a grant check, segregation of duties, and an audit entry per call.

Integration6 min read

MCP-native AI agent governance

MCP lets an agent call any tool a server exposes. Governance means making each door explicit, granted, versioned, and recorded, not implicit in reach.

Integration6 min read

Wrap existing AI agents, do not migrate

Governing AI agents should not mean rebuilding them. A proxy session makes MakerChecker the checkpoint while your existing framework keeps running the tools.

See it for yourself

Reading is one thing. Watch it block an agent.

One command starts the demo: an agent stopped from signing off its own work, and the signed evidence file an inspector can check for themselves.

Book a demo Read the docs

Designed against the rules your auditors already enforce.