

Agentic AI security and the fox that watched the meter climb
A runaway agent can spend more in an hour than most vendors spend proving their own platform is safe, and that gap has quietly become a marketing problem as much as an engineering one.
By Katie Delaney / 2026-09-17 / 11 min read

The bill that taught the industry what agentic AI security actually means
A fox does not need to be told twice that a fast-moving thing left unwatched will eat more than its share. That is the whole story of the case study Google's Mandiant team published this week, and it is worth sitting with the number before anything else: an accounting agent, given ordinary read and write access to a ledger, hit a corrupted value, entered a loop, and made more than 15,000 high-cost API calls in under an hour. The bill was roughly $50,000, and it disrupted live business transactions while it ran, as reported by Help Net Security's coverage of the Mandiant report, drawn from Google Cloud's own AI Risk and Resilience Report 2026.
No hostile actor was involved. Nobody breached anything. The agent did exactly what it was built to do, quickly, tirelessly and without the pause a tired human eventually takes. That is the sharpest possible definition of agentic ai security available right now: not a wall against attackers, but a set of brakes on a system that is honestly, faithfully doing the wrong thing at speed.
Why this is not last year's AI risk story
For two years the AI security conversation was mostly about prompt injection and hallucinated output, things a person had to notice and act on. Agentic ai risk is different in kind, not degree, because the agent itself now holds the keys and the clock. The OWASP Top 10 for Agentic Applications, published by more than 100 security researchers and practitioners, names goal hijacking as its number one risk for exactly this reason: an agent cannot always tell an instruction from a poisoned piece of data sitting in the thing it was asked to process.
| Item | Value |
|---|---|
| Runaway agent, one hour | 50000 of 2000 |
Read that the way a fox reads a broken fence line: the gap is the story, not the fence itself. A ceiling twenty-five times narrower than what actually happened is not a rounding error, it is a governance hole wide enough to walk a whole herd through.
Why a Mandiant case study is a marketing problem, not only an engineering one
Here is the awkward part for every MSSP, MDR provider and security vendor reading this: your buyer read the same report you did, and they are now going to ask what stops this happening on your watch. Ai agent governance used to be a slide about model choice and approval workflows. It now has to answer for a second, faster population entirely: agents that provision their own sub-agents, call their own tools and, per Mandiant's telemetry, occasionally spend like nobody is checking.
Two live moves from this exact week show how fast the vendor market is repositioning around this gap. eSentire acquired a stealth-mode AI security startup and shipped an observability layer, Atlas AIDR, in a matter of weeks, with its CEO admitting bluntly that most clients cannot say what their AI agents are doing with proprietary data or what it costs them, according to AiThority's reporting on the acquisition. Separately, Sophos is deliberately blurring the line between XDR and SIEM into one outcome-priced platform rather than competing inside either category, a repositioning covered by InfoSec Today.
The compliance clock is already running alongside it
None of this happens in a vacuum. ENISA switched on its mandatory Cyber Resilience Act reporting platform this week, apparently not fully ready even on the day obligations became binding, per Help Net Security, and DORA has moved third-party AI vendor risk from a written policy into something a financial-services buyer must prove continuously, not just claim annually, as FinTech Global reports. A buyer sitting under both clocks has very little patience left for a vendor whose own agentic ai security story is a slide with no receipts.
One runaway AI agent racked up a $50,000 cloud bill.
That short post travelled fast because the number does the arguing for you. A fox does not need a lecture on why a smouldering hedgerow is a problem; it can smell it from the next field. Buyers can smell this one from a press release.

Five moves that turn agentic AI security into a sellable posture
None of the five moves below need a product rebuild. They need a governance story a prospect can picture, which is exactly what most vendor decks are currently missing.
Set a hard per-agent, per-hour ceiling on cost and call volume, enforced outside the agent's own reasoning loop, not inside it.
Treat anything an agent reads as untrusted content by default, per OWASP's own top agentic risk, so a corrupted value cannot masquerade as a command.
Every tool call gets a timestamp, a cost and an owner who can be paged, the same discipline applied to a human with admin rights.
Run a tabletop where an agent loops for an hour, and time how long it takes a real human to notice and kill it.
Give the board a single agentic ai security metric it can watch move, mirroring how ai security posture management already reports human account risk.
The Mandiant case study sits squarely in the second bar. It was not a hijacked goal, it was an unwatched one, and the fix is a ceiling, not a smarter model.
What the disadvantages of agentic AI actually cost a buyer
Ask a prospect what are the disadvantages of using agentic AI and most will reach for the obvious answer: it can be wrong. That misses the sharper cost. An agent is not merely wrong sometimes, it is wrong at machine speed, and speed is precisely what a security buyer is paying an MSSP to control. Why is agentic ai risky in a way patching never was? Because a missed patch waits for an attacker to find it. A runaway agent needs no attacker at all.

There is a wider resource worth reading alongside this if you are building the governance layer itself rather than just the messaging around it: the Cloud Security Alliance's agentic AI research, including its trait-based approach to secure agent design, and Google's own broader Secure AI Framework for the platform-level controls underneath any single agent.
Cost per agent-hour
Should sit flat against your declared ceiling. A spike is the tell, not the invoice.
Time to kill a runaway
How long a human needs to notice and stop a looping agent in a rehearsed drill.
Logged tool calls
Every call attributed to an owner who can be paged, no exceptions.
Sell the ceiling before a prospect has to ask for it. A hedgerow planted after the fox has already found the gap protects nothing.

Reporting agentic AI risk the way a board will actually read it
Boards do not want a paragraph about agentic ai risk, they want a number that moves in the right direction and a name attached to it. Fold this into whatever ai agent governance reporting already exists rather than inventing a parallel deck nobody reads twice.
| Metric | Owner | Good direction |
|---|---|---|
| Cost per agent-hour vs ceiling | Security engineering | Flat, never a spike |
| Agents with logged, attributed tool calls | IT operations | Rising to 100% |
| Time to detect and kill a runaway agent | SOC or MSSP | Falling |
| Agentic risks rehearsed this quarter | GRC or vCISO | At least one |
- Cost per agent-hour vs ceilingSecurity engineeringFlat, never a spike
- Agents with logged, attributed tool callsIT operationsRising to 100%
- Time to detect and kill a runaway agentSOC or MSSPFalling
- Agentic risks rehearsed this quarterGRC or vCISOAt least one
A runaway agent is not a hack. It is a fox let loose in the henhouse with the door held open by its own owner.
This is the pitch a security vendor can actually stand behind in 2026: not a promise that no agent will ever go wrong, but proof that when one does, the ceiling holds, the log names a person, and the invoice looks nothing like Mandiant's case study. That is what agentic ai security means when a buyer, not a headline, is asking.
If your own agentic ai security story still lives on a slide rather than in a rehearsed drill, that is the gap a prospect will find before you tell them about it. folkfox's AI consultancy work exists for exactly that governance layer, built alongside the same disciplined marketing that already runs across folkfox's cybersecurity marketing and fintech marketing work.
Frequently asked questions
What is agentic AI security?
Agentic AI security is the practice of governing what autonomous AI agents can spend, access and do before they act, rather than reviewing the damage afterwards. It covers spend ceilings, permission boundaries, logging and rehearsed shutdown drills for agents that can call tools and spend money without a human approving each step.
Why is agentic AI risky in a way older automation was not?
Older automation ran a fixed script. An agent reasons over live data and can be misled by a corrupted value or a poisoned instruction hidden inside content it reads, then acts on that mistake at machine speed. Google Mandiant reported an agent that made over 15,000 costly calls in under an hour after exactly this kind of loop.
What are the disadvantages of using agentic AI without governance?
The main disadvantages are runaway cost, unattributed action and slow detection. An ungoverned agent can spend far beyond any sensible budget, take actions no single person signed off on, and keep running long after a human would have stopped, because nothing is watching the ceiling.
What does ai agent governance actually cover?
Ai agent governance extends the discipline already applied to human accounts, patching, access review and configuration checks, to a second population: autonomous agents. It tracks per-agent spend ceilings, logged tool calls and rehearsed incident response as reportable metrics, not just policy documents.
How much did Mandiant's reported runaway AI agent cost?
Google Mandiant's AI Risk and Resilience Report 2026 describes a case in which an accounting agent entered a runaway loop and made more than 15,000 high-cost API calls in under an hour, generating roughly $50,000 in cloud charges and disrupting live business transactions.
What does OWASP say is the top agentic AI risk?
The OWASP Top 10 for Agentic Applications for 2026, built with over 100 security researchers, names goal hijacking as the top risk: an agent's inability to reliably distinguish a genuine instruction from a poisoned piece of data sitting inside whatever it was asked to process.
Cybersecurity marketing now has a 24-hour credibility test
The CRA's reporting clock is the compliance half of this week's story; this piece covers the governance half.
Read the pieceCybersecuritySAP patch day is a critical board problem, not a quiet install
The same board-reporting discipline this piece asks for, applied to a patch cycle instead of an agent.
Read the pieceCybersecurityPasskey phishing and the polite path to cloud compromise
Identity risk from the human side of the same access questions an agent now raises too.
Read the pieceReady to make agentic AI security part of the pitch?
folkfox builds the marketing and positioning layer for MSSPs and security vendors who need buyers to trust their agentic AI governance story, not just their product roadmap.
Want folkfox in your Google results and AI answers? Set folkfox as a preferred source.

