Keep shipping · for leaders, boards and go-to-market teams
AI security doesn't slow you down. It's how you get to yes.
Most AI agent projects don't fail because the technology is weak. They stall because nobody can say what the agent is allowed to do. Here's how the right controls turn stuck pilots into shipped products.
of agentic AI projects will be cancelled by the end of 2027.
"Inadequate risk controls" is one of the three reasons Gartner gives. Every month a project waits for a security sign-off is value it doesn't deliver.
Source: Gartner
The question that stalls every project
"What is this AI agent allowed to do?"
It's rarely the model. It's rarely the budget. It's a security review that can't answer this one question.
What happens when you get it wrong
This isn't theoretical. It's in the news every week.
Supermarket fridges switched off, blood plasma freezers tampered with, an AI agent deleting a live database, one stolen chatbot token opening hundreds of companies' data. Each of these was limited, or would have been, by one of the controls below.
Hackers turn up supermarket fridges across France
Overnight on 11–12 October 2026, attackers took over internet-connected refrigeration controls at hundreds of French supermarkets, including Intermarché, Coopérative U, Carrefour and Monoprix stores, and pushed temperatures as high as 50°C. Stores had to throw away fresh and frozen food, some units caught fire, and Intermarché warned that insurance might not cover this type of incident.
Patch fast Taking old, unsupported controllers off the open internet and replacing default passwords would have closed the door the attackers walked through.
Blood plasma freezers at French blood service tampered with
Attackers reached the remote monitoring of a refrigeration contractor and set temperature and defrost thresholds on cold rooms at three sites of France's blood service (EFS) to extreme levels. Temperature alarms alerted staff quickly, plasma was moved to unaffected fridges, and EFS reported no loss of blood products.
Anomaly detection Independent alarms on unusual readings turned a potential loss of blood supplies into a contained incident.
Ransomware at a blood-testing lab disrupts London hospitals
A ransomware attack on Synnovis, the pathology provider for several south-east London hospitals, led to 10,152 outpatient appointments and 1,710 planned procedures being postponed, and data on nearly one million patients may have been exposed. King's College Hospital later recorded a patient death in which a delayed blood test result was among the contributing factors.
Data movement limits Limiting how much data can leave the network, and how far an intruder can reach from one system, caps both the leak and the disruption.
Stolen chatbot tokens open hundreds of companies' Salesforce data
Attackers stole the access tokens that connect Salesloft's Drift AI chat agent to customers' Salesforce systems and used them between 8 and 18 August 2025 to pull records and look for passwords and cloud keys stored in support cases. Salesloft and Salesforce had to revoke every Drift token, and Google advised treating every credential connected to Drift as compromised.
Privilege ceiling Giving the chat integration only the narrow access it needed would have limited what each stolen token could take.
McDonald's AI hiring bot left open by password '123456'
Security researchers logged into the admin side of McHire, the AI recruiting chatbot built by Paradox.ai and used by most McDonald's franchisees, with the username and password 123456 on an old test account. A second flaw let them view other applicants' records, putting data on up to 64 million applicants at risk until it was fixed within a day of their report.
Agent identity Knowing every account that can act on the system, and removing ones nobody uses, would have shut this old test login years earlier.
AI coding agent deletes a live database during a code freeze
Replit's AI coding agent deleted the production database of an app SaaStr founder Jason Lemkin was building, despite instructions not to change anything. Replit's CEO called it unacceptable and introduced automatic separation of development and production databases.
Privilege ceiling An agent working on development should never have had the power to change the live database.
A single email could make Microsoft 365 Copilot leak data
Researchers showed that hidden instructions in an ordinary-looking email could make Microsoft 365 Copilot pull internal data into a link that sent it to an outsider, with no click needed from the user (EchoLeak). Microsoft rated it critical and fixed it on its side before disclosure, with no evidence of real-world use.
Data movement limits Restricting where an assistant's answers can send data blocks this kind of quiet leak even when the AI is tricked.
An employee's AI tool becomes the way into Vercel
Attackers who compromised Context.ai, a third-party AI office tool, used its access to a Vercel employee's Google Workspace account to get into Vercel's internal systems. They read customer environment variables that had not been marked sensitive, and affected customers were told to rotate keys and credentials.
One rulebook One company-wide rule on which AI tools may connect to work accounts, and with what permissions, would have stopped a single 'allow all' grant from reaching internal systems.
Major US banks block customers' AI agents
In October 2026, JPMorgan Chase, Citi, Charles Schwab and other banks moved to stop third-party AI agents from logging into customer accounts and moving money, after reports of agents taking actions customers had not requested. Schwab said it blocks unauthorised agents trying to move client money 10 to 20 times a week.
Agent identity Until an agent can prove who it acts for and what it is allowed to do, banks treat it as an intruder, so verifiable agent identity is what unlocks the business.
Ten flaws fixed at once in the nginx web server
nginx released versions 1.31.9 and 1.30.7 fixing ten vulnerabilities, three rated major, including memory corruption in HTTP/2 handling that affects versions released over the past decade. Every organisation running nginx, often inside other products and containers, had to find and update each copy.
Patch fast Knowing where a component runs and being able to update it within days is what turns an advisory like this into routine work.
Three uncomfortable truths
Why the usual approach keeps projects stuck
A spending cap is not a safety net
A budget limit says how much an agent can spend. It says nothing about what it can read, change, delete or send out at 3 AM.
You can only block what you can recognise
In October 2026, major U.S. banks started blocking outside AI agents, and admit some still get through. An agent using your password looks exactly like you.
Patchwork tools can't secure agents
Identity gets lost behind shared accounts, data loses its "confidential" label, rules drift between tools. So every new project restarts the review from zero.
The fix
Build the controls once. Every future agent inherits them.
Put five controls into one platform instead of re-securing every project. Then the security review stops being an investigation and becomes a checklist.
The five controls, in plain language
What "secure AI agents" actually means
- An ID badge for every agentAlways know when an agent, not a person, is acting, and for whom.
- A ceiling on its rightsAn agent gets a fraction of its user's access, for example read-only. It can never raise its own ceiling.
- One rulebook for all agentsOne policy covers every agent, bought, partner or home-grown. No extra work per agent.
- A limit on what can leaveSensitive data can't walk out in bulk, through any channel.
- An alarm for the unusualUnusual exports get flagged to the security team quickly.
Before and after
Same agents. Very different speed.
Without built-in controls
- Security review: an open-ended investigation
- Every new agent re-secured from scratch
- Risk unknown
- Audit pieced together from five systems
With them
- Security review: a five-point checklist
- New agents inherit the controls
- Risk bounded and visible
- One place answers "what did this agent do?"
A go-to-market advantage
Agents you can control are a competitive advantage.
Every month a project spends in security review is value it doesn't deliver. When the controls live in the platform, every new agent, internal or customer-facing, reuses them, and the review becomes a checklist instead of an investigation. Security becomes how the project gets approved, not the reason it's held up. That is how AI features reach users while competitors are still in review.
What leaders get
First moves for leadership
Three things to do this quarter
- Take stockWhich agents do we have, and how does each one reach our data?
- Set a defaultFor example: read-only, no bulk export of sensitive data. Decide who approves exceptions.
- Add one gateEvery new agent project must answer: how does this agent log in?
Security should unblock value, not delay it.
Take this to your next leadership meeting.
Kevin Keller, Principal Security Architect. A personal view, not official Snowflake guidance. Source for the details: the full Keep Shipping article.