The takeaway
AI labs are building government-style insider-risk teams, but the threat they are hiring against exists at every company running agents with broad credentials. Scoped keys, tool-call logs and approval gates are the minimum.
Why it matters for builders
Insider-risk controls now matter for agent stacks: give each agent a scoped credential, log tool calls rather than chat text, and keep a human approval gate on money, deletions and customer messaging.
AI labs triple counterintelligence hiring to guard trade secrets
Anthropic hired 33 people into counterintelligence roles in 2026, up from 14 last year and nine in 2024. OpenAI went from nine to 22. Nvidia added 16. Most postings pay between $200,000 and $500,000, according to Revelio Labs data reported by Fortune and the original Reuters reporting, and the labs are still hiring.
What happened
Counterintelligence barely existed at AI labs three years ago. Now it is an org-chart line item. Anthropic put former CIA deputy director David Cohen and former NSA cybersecurity director Dave Luber on its national security advisory council. OpenAI named former NSA director Paul Nakasone to its board in 2024.
The hiring followed the same logic. OpenAI is recruiting a senior global security investigator with counterintelligence as a core skill, Anthropic is advertising for a threat-intelligence security engineer, and ASML wants a Taiwan-based intelligence officer.
Nvidia's board is investigating whether company secrets were exposed in the relationship between its former CEO and an executive at Megaspeed, a firm implicated in chip diversion. A month earlier, OpenAI fired an engineer accused of planting a recording device in an executive conference room.
"The AI firms have snapped up a lot of security talent," Christopher Falkenberg, a former Secret Service agent now at Insite Risk Management, told Reuters. "Demand outstrips the supply of people with relevant experience."
Former FBI counterintelligence operative Eric O'Neill told Fortune the approach rarely starts with a request for secrets. It starts at a conference, over drinks, on a dating app, and it looks like a relationship moving unusually fast before anyone asks about a roadmap.

Why it matters
Frontier labs protect weights, training recipes, customer contracts and roadmaps. Those assets live in systems that employees, and increasingly agents, can read.
The uncomfortable lesson for smaller teams is that the threat model which justifies a $500,000 specialist does not disappear at five people. It just gets cheaper to exploit. If a company with dedicated insider-risk staff still needs a board investigation after one relationship goes wrong, a team running agent workflows on a shared admin key has no controls at all.
Agents make it worse. An agent with a database credential, a billing account and an outbound messaging tool is a person-shaped permission set with no loyalty and no memory of what it should not do.
What builders should copy
Three controls show up repeatedly in how these companies structure access, and none need a security budget.
Give every agent its own scoped credential instead of borrowing a human admin key. Log the tool calls, not just the conversations, because the trail that matters is which system was touched and by whom. And put a human approval gate in front of anything that can move money, delete records or message a customer.
None of that stops a determined adversary with a fake identity and a conference badge. It does mean that when something goes wrong, you can answer the only question that matters: what exactly did it touch.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
15 November 2026
15 November 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.


