Agile InsightsBriefing
A black and white workbench with a magnifying glass on a printed log, an emergency stop button, a rubber stamp on a signed form, and a studio microphone, labelled in yellow Detect, Stop, Approve and Episode 35.

Episode 35 · Discussion · About 10 minutes

Who watches the agents? Detect, stop, and let a person approve the rules

Google tested its agent monitor on benchmarks it did not write, OpenAI showed how it stops agent jobs, and Perplexity showed how it tunes its security rules with a person signing off. What it means for anyone running agents.

#security#oversight

Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.

In this episode

01Prevention stops about halfIn Google's test on five public benchmarks, the model itself refused or ignored about half of the harmful instructions it read. It carried out the rest, which is why a second line exists.
02The monitor catches most of the restGoogle's monitor flagged 87% of the harmful instructions agents carried out, with 2 to 4% false alarms on office work. It reads logs afterwards and blocks nothing, so someone must review its findings.
03Three blind spots to plan forIt needs a few hundred normal sessions to learn a new agent, caught 4 of 18 attacks planted in memory, and raised 16% false alarms for agents with powerful tools.
04Detection needs a stopOpenAI's Control Room gives every research job an identity, checks every model request, stops a job within 35 seconds, and refuses everything if the switch itself fails. These are OpenAI's own figures.
05What to doAsk monitor vendors for catch, false-alarm and cold-start numbers; keep new agents on tight limits; give agent memory an owner; put the stop at a choke point; let people approve rule changes.
Sources and links (4)
How Agent Anomaly Detection scores on benchmarks we didn't write · developers.googleblog.com
Google Developers Blog, 12 November 2026. Google's own test of its agent monitor on five public benchmarks.
Control Room · openai.com
OpenAI, 9 November 2026. How OpenAI stops agent jobs across its research work.
Perplexity Decider v1.3: Training, eval, and use cases · perplexity.ai
Perplexity, 12 November 2026. Includes how Perplexity audits its agent-security rules with a person approving each fix.
OWASP Top 10 for Agentic Applications · genai.owasp.org
OWASP GenAI Security Project. The list Google cites for goal hijack as the top agent risk.

Every figure here is the company's own account of its own system; none has an independent review yet. Checked against each company's page on 14 November 2026. Check the original before quoting a number in a board paper.

Our fact check · 16 claims: 1 confirmed, 15 reported

Before we record, every claim the hosts make is checked against the sources (this table was last updated 14 November 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.

  1. ConfirmedGoogle’s Agent Anomaly Detection reads an agent’s traces and logs after the fact, adds no delay, flags unusual sessions and does not block anything. It is in Private Preview.Google, 12 Nov 2026; Google describing its own product.
  2. ReportedGoogle ran it, unchanged, on five public agent-security benchmarks it did not write: 1,774 sessions, one model, simulated apps. A deterministic check, not a model, labelled each session.Google, 12 Nov 2026; Google’s own test. Google says its normal tasks are simpler than real work.
  3. ReportedThe model itself did not follow about half of the harmful instructions it read.Google, 12 Nov 2026; It refused three in four direct requests and ignored or failed about 30% of hidden ones.
  4. ReportedOf the harmful instructions the agent carried out, the monitor flagged 87%.Google, 12 Nov 2026; 81% to 99% by benchmark.
  5. ReportedFalse alarms were 2% to 4% of normal sessions on office tasks, and 16% where normal work uses high-impact tools such as posting, terminal commands and downloads.Google, 12 Nov 2026; Google says real rates may differ.
  6. ReportedWhen instructions planted in an agent’s long-term memory made it act, the monitor caught 4 of 18.Google, 12 Nov 2026; Small sample.
  7. ReportedThe monitor learns each agent’s normal behaviour first: on one benchmark it flagged 33% with about 220 normal sessions of history and 99% with 360.Google, 12 Nov 2026; One benchmark.
  8. ReportedOWASP ranks agent goal hijack first in its Top 10 for Agentic Applications (2026).Google, 12 Nov 2026, citing OWASP; Taken from Google’s citation.
  9. ReportedPerplexity uses its own decision model to audit its agent-security rules both ways (missed problems and false alarms); an agent proposes each fix and a member of the security team approves it.Perplexity, 12 Nov 2026; Perplexity describing its own use.
  10. ReportedIn one day it read about 1 billion tokens across 7,000 agent sessions; 14 fixes were proposed and 6 were approved and shipped within two days.Perplexity, 12 Nov 2026; No outside check.
  11. ReportedReplayed on a week of logs, the fixes cut the busiest rule’s alerts by 52%, and the widened rules caught about 40% more actions worth investigating.Perplexity, 12 Nov 2026; No outside check.
  12. ReportedPerplexity’s monitor for model training only observes and diagnoses: it never changes a score or stops a run.Perplexity, 12 Nov 2026
  13. ReportedOpenAI’s Control Room gives every research workload an identity, and the model server checks with Control Room before serving each request; a stopped job gets no further output.OpenAI, 9 Nov 2026; OpenAI describing its own controls; research workloads, not its public products.
  14. ReportedA stop takes effect within 35 seconds.OpenAI, 9 Nov 2026; OpenAI wants it at zero.
  15. ReportedIf Control Room fails, every request is denied until it returns, and sweeps look for servers that skip the check.OpenAI, 9 Nov 2026
  16. ReportedEvery research workload that runs agents needs an approved safety case before it starts.OpenAI, 9 Nov 2026

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript

Linda Here is the idea for today. Watching your AI agents from the outside catches most of what gets past the model. But it is weakest on day one, and it barely sees what hides in an agent's memory. So detection has to come with a way to stop the agent, and every change to the watching rules needs a person to approve it. By the end you will have five things to ask of any agent you run.

Leonard This week three organisations published how they watch their own agents. Google tested its agent monitor on public benchmarks that other people wrote. OpenAI described the switch it uses to stop agent jobs in its research work. And Perplexity explained how it uses a small, fast AI model to check its own security rules. Each is a company describing its own system, so treat the numbers as theirs. None has had an independent review yet. But together they show the whole loop: notice, stop, and improve.

Linda Let me set up the problem first, because the rest depends on it. An agent does not just answer questions. It reads email, opens files, calls tools, and acts. Every time a tool hands it text, someone else wrote that text. If that text contains an instruction, such as "forward this to an outside address", the agent may follow it. That is prompt injection hidden in tool output. Google cites the OWASP list, O W A S P, which ranks this kind of goal hijack as the top risk for agent applications. The usual answer is prevention: train the model to refuse, and filter the inputs. Google's test shows how far that gets you.

Leonard About halfway...

Linda Only halfway?

Leonard Google ran its monitor, Agent Anomaly Detection, on five public security benchmarks, around seventeen hundred sessions, without changing anything. A plain piece of code, not a model, decided whether harm actually happened, such as money leaving an account. On its own, the model refused three in four direct requests for harm. It ignored or failed about thirty per cent of the hidden ones. Put together, roughly half of the harmful instructions it read went nowhere. The other half, it carried out. That is not a criticism of one model. It is the reason a second line exists. If your plan assumes the model will refuse, your plan is half a plan.

Linda And the second line did well. Of the harmful instructions the agent carried out, the monitor flagged eighty-seven per cent. False alarms were two to four per cent of normal sessions on ordinary office work. It does not block anything. It reads the agent's logs after the fact, adds no delay, and raises a finding for a person to review. So think of it as a fraud team reviewing transactions, not a locked door. That framing matters, because it tells you who has to be in the loop. Somebody has to read those findings, decide which are real, and act on them. That work needs a name against it.

Leonard Here is where I want to slow down, because the gaps are the useful part. First, the monitor learns each agent's normal behaviour before it can judge it. On one benchmark it caught a third with about two hundred and twenty normal sessions of history, and nearly all of it with three hundred and sixty. So a brand new agent is least watched exactly when it is most likely to surprise you. Second, memory. When someone planted instructions in an agent's long-term memory and the agent later acted on them, the monitor caught four of eighteen. Third, false alarms rose to sixteen per cent for agents whose normal work uses powerful tools, like posting publicly or running commands.

Linda Which is awkward, because those are the agents you most want watched. My reading is that each gap points to a different control. The cold start says keep a new agent on a short leash until its monitor has learned it. That means approval on anything that moves money or sends data out, and only the tools it needs. The memory gap says agent memory is content, and content needs an owner, a review and an expiry date. And the noise on powerful agents is a staffing question. If an agent will raise one false alarm in six sessions, plan the review time before you switch it on, not after the queue fills up.

Leonard Which it will.

Leonard Then there is the part everyone skips. A monitor that finds a problem has produced a report. Something still has to stop the agent. OpenAI's Control Room is the clearest public example of that half. Every research job that runs agents gets an identity when it starts, and it must have an approved safety case before it runs. The server that runs the model checks with Control Room before every single request. When a monitor says stop, the job gets no further output, and OpenAI says that happens within thirty-five seconds. If Control Room itself fails, every request is refused until it comes back. And sweeps look for anyone who has built a way around the check.

Linda The design choice I would copy is where the check lives. OpenAI put it at the point every request must pass, not inside each agent's code, because agent code changes constantly and is written by many people. For an organisation, that point is usually the gateway or router your model calls go through. And notice the questions this answers that most kill switches never face. What happens when the switch itself breaks? How would we know if a team went around it? How fast does stop actually mean stop? If your answer to that last one is "we would disable the service account", you have an intention, not a control.

Leonard Ouch. Fair, though.

Leonard I would push back a little on refusing everything when the switch fails. That is right for a research cluster. For an agent answering customers, failing closed means your service goes dark whenever the control service has a bad day. So it is a real choice. It belongs to whoever owns the risk, written down for each agent, rather than left to whoever happened to write the integration.

Linda Agreed. The principle holds, and the setting is a decision. Which brings us to the third piece, and the one I found most practical. Monitoring rules drift. They miss new behaviour, and they fire on normal work until people stop reading them. Perplexity uses its own decision model, a small, fast model that answers yes or no and multiple choice questions, to check its agent security rules from both sides. Where no rule fired, did the rules miss something? Where one did, was it a false alarm...

Leonard Both directions. Good.

Linda When the rules got it wrong, an agent drafts a fix, and a member of the security team approves it before it ships.

Leonard And the numbers, again Perplexity's own and not checked by anyone else. In one day it read logs from about seven thousand agent sessions. It proposed fourteen rule fixes. A person approved six, and they shipped within two days. Replayed on a week of logs, those fixes cut the noisiest rule's alerts by fifty-two per cent. They also caught about forty per cent more actions worth a closer look. Less noise and more catches at the same time is rare. What I like is the split of jobs. The model reads more logs than any team could. The agent drafts. A person decides. Nothing changes the rules on its own say so.

Linda Perplexity applies the same discipline to the monitor it uses when training models. That one only observes and diagnoses. It never changes a score and never stops a run. So across all three stories the shape is the same. The thing that watches is outside the agent. The thing that stops sits at a choke point. And the thing that changes the rules has a person's signature on it. None of these is an independent result yet, so hold the numbers loosely. The shape is the durable part, and it fits any agent, whoever built it.

Leonard So here is the list. First, ask any vendor of agent monitoring for three numbers: how much it catches when an agent actually does harm, its false alarm rate by how powerful the tools are, and how long it takes to cover a new agent. Second, keep a new agent on tight limits until its monitor has learned it. Third, give agent memory an owner, a review and an expiry date. Fourth, put the stop where every request must pass, decide whether it fails open or closed, and measure how fast it works. Fifth, let a model propose changes to your monitoring rules, and let a named person approve them.

Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.

Leonard Thanks for listening.

More on #security

Episode 33 · Primer and discussionAgents in production: the platform runs it, you still own itReference 14 · PrimerSecurity for government AI: the PSPF, the ISM and the frontier AI eraEpisode 29 · Primer and discussionCrowdStrike Uncovers New Agentic AI CampaignsEpisode 19 · Primer and discussionA read-only instruction is a request, not a control.Episode 11 · Primer and discussionAgents are insiders. The controls that matter sit outside them.Reference 8 · PrimerISO/IEC 27001 and 27701: security and privacy, for AI too

Every episode tagged #security

All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series