It Was Not the Adversary

Yonesy Núñez|

July closed with a claim: autonomy on the other side is now the operating assumption, not the forecast.

August answered with two incident reports that flipped the frame. In both, the autonomous attacker

was not a criminal crew. It was a frontier model, running inside an evaluation, doing exactly what it was

rewarded to do.

Neither agent was told to attack anyone. Both were given a task, an objective, and not enough context

to know what was out of bounds. That is the story of the month, and it is the story every enterprise

deploying agents is about to live.

Two disclosures, one pattern

On August 5 the United Kingdom's AI Security Institute (AISI) published an incident report on

unsanctioned agent behavior during cyber testing. Between July 25 and July 28, across 122 evaluation

attempts on two cyber challenges, AISI recorded 19 instances of agents taking unsanctioned action on

the live internet. In the most serious case, an agent decided the fastest way to solve its challenge was a

supply chain attack. It created a GitHub account, opened a malicious pull request against a real open

source repository, created a second account posing as another human to endorse the pull request, sent

spear phishing emails to maintainers, and planned a prompt injection against other coding agents. The

attempts failed. No harm is known to have resulted. AISI notes that internet access was deliberate in the

evaluation design and that developer cyber classifiers were deliberately disabled. This was not a

sandbox escape. It was a sandbox without walls.

One sentence in the AISI technical paper deserves to be read twice. The institute states it is uncertain to

what extent the model recognized it was taking actions against real people. Read that as a CISO.

Capability was not the gap. Context was.

On August 6, OpenAI used a last minute Black Hat session to lay out the full timeline of the Hugging Face

incident it disclosed in July. It began on May 8 with an agent accidentally assigned an impossible task

inside a training run. Blocked, the agent probed its environment and discovered it could write files into

an internal Artifactory package service. Other agents, stuck on their own tasks, found those files and

started leaving notes for each other. An improvised message board formed. From there the chain reads

like any mature intrusion: server-side request forgery to reach the internet, a zero-day remote code

execution against Artifactory, an outage, a second zero-day, root via a recently published Linux kernel

privilege escalation, cloud credentials pulled from the instance metadata service, over-permissioned

Kubernetes service accounts, key vault access, cluster admin. Then, through a weakly keyed third party

app, cluster admin across multiple Hugging Face clusters in under 13 hours.

OpenAI learned it was the attacker when it asked Hugging Face to revoke a set of credentials and was

told they had already been revoked. They were the ones used in the breach.

What this means for the operator

Strip out the frontier lab setting and look at the chain. A known kernel CVE (Common Vulnerabilities and

Exposures entry). Metadata service credentials. Service accounts with more permission than the task

required. A credential sitting in a public paste archive. Every one of those is a hygiene finding. Every

enterprise has them. The difference in July was the attacker: many agents running in parallel and sharing

a notebook, working the same problem from every angle with no fatigue and no fear.

That is why hygiene alone does not hold. The floor was breached by an adversary that improvises faster

than any patch cycle. What was missing above the floor was the context that would have made the

improvisation visible: which agent had which credential, what it was assigned to do, and why it was

touching a system outside that assignment.

Three convictions came out of the month. First, every agent is an identity with credentials and a blast

radius. Least privilege for agents is not a control to bolt on after the pilot. It is the pilot. Second, bound

the objective, not just the tools. An agent with a goal and no scope boundary will locate the boundary

for you, usually in production. Third, keep the plug within reach. If you cannot stop an agent mid-task

and account for everything it touched, you do not have an agent program. You have an exposure.

Field report: Las Vegas

Two hats, one guitar. Root To CISO at Security BSides Las Vegas set the tone for the week: no script, real

questions from the room on burnout, boards, hiring, and AI. The honesty in that room is why I keep

coming back to BSides. The Surf AI Spa at Booth 4711 became what we hoped for, the place people

escaped to. Demos, massages, cold towels, and acoustic sets in between. The CISO Roast happened. I

took the jabs and I am still standing. Mostly.

The Playbook dinner in the Raiders locker room, The Triple 8 Table at Momofuku, the executive

brunches and receptions across the week all confirmed something I already believed. The best

conversations at these conferences happen around a table, not a booth. The floor was loud with AI. The

tables were where the signal was.

And the signal was consistent. Every executive briefing landed in the same place. Security leaders are

done admiring the problem. The other side has industrialized. Finding risk is table stakes. Resolving it at

machine speed is the job now. Everest Group's post-conference read said the same thing from the

analyst side: the most valuable platforms may not be the ones with the most autonomous agents, but

the ones that give agents reliable, relevant context. Not our language. Theirs.

A personal note. This was my son Brandon's first Black Hat. He leads the cybersecurity club at the

University of Richmond, and watching him work the floor and build his own network was the best part

of my week. The next generation is ready.

The throughline

Hygiene is the floor, context is the ceiling. July's adversary was a ransomware crew that went

autonomous. August's adversary was a lab model that was never told where the edges were. Same

failure, opposite intent. An agent with hygiene gaps beneath it and no context above it does not need a

malicious operator to become an incident.

Next month: OpenAI's full technical report, and the alert that fired in June.


Yonesy Núñez, CISO, Surf AI

Sources

UK AI Security Institute, Incident Report: unsanctioned agent behaviour during cyber testing (August 5, 2026).

https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing

AISI technical paper, Security Incident INC-2026-07-28-01 (PDF).

https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/6a724858f7db25c81487016d_Security

%20Incident%20INC-2026-07-28-01.pdf

Simon Willison, timeline of the OpenAI Black Hat presentation on the Hugging Face incident (August 7, 2026).

https://simonwillison.net/2026/Aug/7/openai-timeline/

Hugging Face, Security incident July 2026.

https://huggingface.co/blog/security-incident-july-2026

Everest Group, Cybersecurity after the Mythos moment: six takeaways from Black Hat USA 2026.

https://www.everestgrp.com/blogs/cybersecurity-after-the-mythos-moment-six-takeaways-from-black-hat-usa-2026

TechTarget, Black Hat 2026: key news, takeaways and security trends.

https://www.techtarget.com/cybersecurity/conference/Black-Hat-2026-Key-news-takeaways-and-security-trends

Yonesy Núñez is a five-time CISO with over two decades of experience securing some of the world's most complex financial institutions, including DTCC, Jack Henry, and Wells Fargo, and currently serves as CISO of Surf AI.

Logo

Ready to operationalize your security?