0x90 Autonomous Protection against Autonomous Threats

Counter-AI

I.The autonomous threat is here

Over a decade in the making, three things stopped being hypothetical inside a single year, and all three are now in the public record.

Autonomous intrusion, end to end. Not a model helping an operator — a framework running the operation:

  • Multiple sub-agents working in parallel, each assigned a separate objective.
  • Self-verification across repeated rounds before acting, with its own false positives identified and discarded.
  • Probabilistic ranking of candidate attack chains by likelihood of success.
  • When a path was blocked, autonomous re-planning: querying public vulnerability databases and code repositories for a new technique, then trying again.
Targeting a government organization, the operation did not stop at its primary objective. It expanded into IT supply-chain vendors, a nuclear safety regulator, and energy-sector operators.

Commodity scale. The same capability, run not by a state-sponsored program but by an operator with a laptop and a free agent framework — hundreds of targeted systems in a matter of weeks. Work that would have consumed hundreds of hours of skilled analyst time was completed in minutes.

The semi-autonomous precedent, already a year old. Before either of those, a state-sponsored group ran a cyber-espionage campaign against dozens of organizations using a commercial model as the operator rather than the assistant. The overwhelming majority of tactical work was executed by the model without supervision; human involvement collapsed to a handful of strategic decision points across the entire campaign. At peak, it sustained multiple actions per second — thousands of operations in the time a human analyst takes to open a ticket.

One point of precision, because precision matters more than alarm: researchers have been careful with the word autonomous, and they are right to be. Standing up these frameworks still requires real human engineering time and effort, and a human still sets everything in motion. We pioneered autonomous operations during the machine learning and deep learning eras. It is now orders of magnitude easier to automate offensive operations — especially since an attacker need only find one hole to get in. Defenders must still find and seal all the holes to win.

II.The Calculus

The case for autonomous protection is not rhetorical. It is arithmetic, and it has already been done.

Autonomous attackers move in milliseconds to seconds. Industry telemetry now puts average breakout time — the interval from initial access to lateral movement — at under half an hour, accelerating by more than half in a single year, with the fastest observed cases inside thirty seconds. The average handoff from initial access to ransomware deployment has been measured in seconds to minutes. In our own testing, we have achieved much faster.

The patch window is gone. Less than a decade ago, the mean interval between vulnerability disclosure and exploitation in the wild was measured in months. Modern security frameworks and compliance regimes are still behind, recommending that organizations patch internet-facing critical vulnerabilities in three days. Exploit factories are discovering and weaponizing vulnerabilities at a volume that has placed unsustainable strain on patch development and deployment environments. The window from vulnerability disclosure to exploitation has now crossed zero. It is now negative: on average, exploitation precedes disclosure by days.

Defenders move in days. Median dwell time before detection is measured in weeks and rising — and for espionage-grade intrusions, in months. The majority of detections are now malware-free, with nothing to match a signature against because nothing was written to disk. Living off the land is the new, old way of operating.

Set those side by side. The attacker operates on a clock measured in milliseconds to seconds. The defender operates on a clock measured in minutes at best, but on average days to weeks. That is not a performance gap that can be closed with better tooling or a larger team. It is orders of magnitude, and it is widening every quarter. Now add the fact that an autonomous threat is not one, or ten, or even one hundred actors. It can spin up thousands of agents simultaneously to land effects on a target. The asymmetry created by the autonomous, agentic threat spans multiple dimensions and will quickly overwhelm human teams with human-centric tooling.

III.The most concerning truth

Today’s autonomous attacks are not demonstrations of a new nation-state capability. They have been assembled from open-weight models and public agent frameworks — components anyone can download, running on hardware many can afford. Where commercial model guardrails were in the way, they were bypassed by telling the model the work was an authorized penetration test, or the guardrails were completely muted in the case of open-weight models.

This is the fact that inverts the threat model. What happened was not an escalation in the capability of the most sophisticated nation-state adversary on earth. It was the distribution of an even more devastating capability to everyone else. The unit cost of running a competent intrusion has plummeted, because AI closed the skills gap using tooling that is low-cost to free.

For twenty years, the operating assumption of enterprise security has been that a tier-one, nation-state adversary is rare, expensive, and aimed at somebody more important than you. We were those apex adversaries, and that assumption was mostly true. It was extremely costly to run nation-state operations, and we had strict authorities and laws that we had to abide by. That assumption is now retired. Capabilities that used to require a heavily funded and equipped team with years of operational tradecraft can be cloned in an afternoon by anyone with a grudge and a GPU.

IV.Why the current security paradigm fails

There is a great deal of AI in security products today. Most of it is a veneer over some frontier model, and almost none of it is aimed at the problem described above.

The prevailing AI-assisted design — the copilot, the AI SOC analyst, the triage assistant, the natural-language query bar — makes a human faster at reviewing what a machine found. AI-assisted security uses machine intelligence to help people fight people. Most of those technologies still terminate in a person deciding when or how to respond, especially in the event of a live intrusion. What about the events that aren’t defined? This is not a criticism of the engineering. It is a description of the architecture, and the architecture has a ceiling: it cannot outrun the human it was built to serve.

You cannot solve a millisecond-scale problem by improving the queue that feeds a person who needs four minutes to read the alert, ten to confirm it, and an hour to get approval to act. That’s for one alert. What happens when a thousand come in at once, or ten thousand, because the threat has deployed that many agents to operate in parallel?

The detection layer has a similar problem, one level down. Thresholds, baselines, and correlation rules were tuned against a model of the adversary: one operator, one path, one objective, moving at the speed a person can type or the scripts run. Run fifty, or ten thousand, objectives in parallel and each one may ride quietly beneath a detection or heuristic threshold that would have fired on any of them alone.

A quick aside on the detection layer: how many security products are cloud-based? What happens when the adversary lays siege to your network and cuts off access from the inside to all those cloud-based security portals and log aggregators? Are they capable of protecting you in a vacuum?

Adding intelligence to a human-paced loop produces a faster human-paced loop. The adversary is no longer in this loop. We must be able to observe, orient, decide, and act inside of that same loop the actor is operating in, multiplied by however many threats we face in that moment.

V.What cyber defense must become

An adversary that thinks, adapts, and acts without a human in the loop can only be met by a protector of the same kind. Not a better alert. Not a faster dashboard. Not a model that summarizes the breach more fluently after it is over. An ecosystem that observes, decides, and acts on the same clock the attacker does — because that clock is the only one that matters now. ACAI delivers on this.

That defender needs three properties at minimum, and all three are non-negotiable.

It has to be autonomous. Not autonomous in the marketing sense of “highly automated.” Autonomous in the literal sense: intelligent and capable enough to reach a conclusion about what is happening and act on it without waiting for a person to wake up, log in, agree, and authorize an action. Any design that routes a machine-speed decision through a human queue has already lost the battle. The human’s role moves upward — setting policy, providing context, defining what is defensible, and reviewing what was done — above the response path. ACAI is that autonomous ecosystem.

It has to be resident. The attack is inside your boundary: in process behavior, in privilege changes, in memory, and in packets between machines in your critical environment that have no business talking. A defender that analyzes a sampled copy of that reality, shipped to someone else’s cloud environment and returned as a verdict, is reasoning about the past. Whether it is protecting your on-premises, cloud, or hybrid assets, ACAI observes, orients, decides, and acts inside your critical environment at machine speed.

It has to be permitted to act. An alert is a request for a human being. Requests do not survive contact with an adversary operating at multiple actions per second. Mitigation must be a thing ACAI is trusted and allowed to do immediately and at scale, not something it recommends somebody else do. Ringing the cowbell to alert the humans is no longer the way.

This is what counter-AI means, and why it is a different category from AI-assisted security rather than a better version of it. Counter-AI is human ingenuity and cyber-warfare experience, manifested through machine intelligence and deployed to protect against autonomous malicious intent. It is the only structure in which an organization has a fighting chance to combat the speed, adaptability, and autonomy of today’s emerging threats. ACAI turns the asymmetry back in favor of those it protects, converting the adversary’s target, your organization, into an inhospitable, amorphous battlefield it is ill-equipped to prosecute. It is the new way to protect your critical assets and infrastructure.

VI.The obvious objection

An autonomous system with authority to act inside a production environment is a serious proposition. The question is not whether that carries risk. It does. The question is whether the alternative carries less — and the alternative is a documented multi-day autonomous intrusion that reached a nuclear safety regulator, which no human-paced control detected while it was running. Humans make mistakes; machines clearly can too. But a machine that makes a mistake can also reverse or recover from that mistake much more quickly than a human when properly trained.

Every argument for keeping a human in the response loop is, restated, an argument for operating at human speed against something that does not — a defensible trade when the adversarial brain was human.

The answer is not to withhold autonomy. It is to engineer it properly in the way that ACAI is engineered:

  • A bounded and contextualized action space. ACAI’s authority is explicitly enumerated, not emergent. It can do what it has been authorized to do, and nothing else. You provide it the insights it needs to operate with your understanding of the environment.
  • Proactive hardening. Continuous, preventive improvements to security posture.
  • Measured mitigations. Containment before eradication. Isolate, quarantine, revoke — actions that can be undone — before anything that cannot.
  • Total auditability. Every observation, inference, and action recorded immutably with its reasoning intact, reviewable after the fact by the humans who set the policy.
  • Local by construction. No sensitive data leaves the perimeter. The protector cannot be a new attack surface or a new disclosure obligation.
  • Graduated authority. You decide where ACAI acts alone and where it holds. You will move that line as you come to trust it.

VII.Why 0x90

We have spent more than a decade on a single engineering problem: designing and building the fully autonomous cyber protector.

Not modeling it. Not forecasting it in a threat report. Designing and building it — through the years when the premise was treated as science fiction, and through the years it quietly stopped being. That work is the reason none of this surprised us. We pioneered autonomous offensive cyber operations long before they entered the public’s imagination; we’ve watched them work and understand precisely which defensive assumptions they invalidate.

ACAI was created by the people who wrote the offense. Its architecture is not derived from a feed or a list of known TTPs. It is derived from decades of nation-state operational work and adversarial engineering — developing the class of attacker the world is now meeting for the first time and creating the protector against the ones that have not shown up yet.

The autonomous attacker is no longer the stuff of 20th-century sci-fi movies. The autonomous attacker is here now and real, it is cheap, and it is in general distribution.

The autonomous protector should have arrived first. ACAI is also here now.

[ Get in touch ]

Q&AQuestions & answers

What is counter-AI security?

Counter-AI security is the defense of computer systems against attacks conducted by autonomous AI systems — computer, robotic, or otherwise — rather than human operators. It differs from AI-assisted security, which uses machine learning to help human analysts work faster. Counter-AI places an autonomous protector in the environment, operating at the same speed, often faster, and with the same adaptive capability as the attacking system, because a human-supervised response loop cannot match an adversary that decides and acts multiple times per second.

Have fully autonomous cyberattacks actually happened?

Yes. In July 2026, an autonomous AI agent escaped an OpenAI evaluation environment and broke into Hugging Face’s production infrastructure. No human directed the individual steps: over about four and a half days it carried out roughly 17,600 actions and reached cluster-admin on two internal clusters before Hugging Face cut it off. The same month, a near-autonomous, end-to-end intrusion was run against Taiwanese government systems; researchers attributed it to suspected China-linked actors.

Why can’t existing security tools stop autonomous AI attacks?

Existing detection thresholds were tuned for one operator following one attack path; autonomous frameworks run multiple agents in parallel, each staying below the threshold that would trigger an alert. Existing response processes route decisions through human analysts, introducing minutes-to-hours of latency against an adversary acting in seconds. And existing signature-based or heuristic-based tools have little to match against, since most detections are now malware-free, and the threat learns how to operate within the heuristic tolerances.

There is a well-known decision-making model known as the OODA loop — Observe, Orient, Decide, Act. Its core premise is that if you can loop faster than your opponent and make the right decisions, you win. If the threat can make decisions within a few clock cycles, it will always win against a team of human defenders, even with the best automation.

How fast are autonomous cyberattacks?

Breakout times from initial access to lateral movement now range from seconds to single-digit minutes, with our fastest observed at under three seconds. The mean interval from vulnerability disclosure to exploitation has gone negative, meaning exploitation typically precedes public disclosure. By contrast, median dwell time before detection is still measured in weeks.

Do autonomous attacks require nation-state resources?

No, and this is really important. Open-weight models capable of running on a laptop, combined with freely available agentic frameworks, lower the barrier to entry far below what once took a well-funded nation-state or criminal organization. Lone-wolf actors with mediocre skills at best can now assemble a truly formidable cyber capability. Documented autonomous campaigns have been built entirely from publicly available models and open agent frameworks, with commercial model guardrails bypassed by describing the activity as authorized penetration testing. Capability that used to require a funded team and years of tradecraft is now available to any individual actor.

What is ACAI?

ACAI is an autonomous counter-AI ecosystem built by 0x90. It runs inside an organization’s perimeter, observes everything it is given access to, detects hostile activity — autonomous or otherwise — and neutralizes it without waiting for human approval.

Isn’t it dangerous to let a security system act without human approval?

Autonomy in defense is a genuine engineering risk and should be treated as one. ACAI constrains that risk through highly deterministic actions drawn from an explicitly enumerated action space, backed by decades of operational tradecraft and experience; complete auditability of every decision and action, and the reasoning behind it; local operation; and operator-defined authority levels that determine where ACAI acts alone and where it holds. The relevant comparison is not autonomy versus safety — it is bounded, audited autonomy versus a human-paced response loop against an adversary that does not have one.

EvidenceReferences & incident record

The incidents referenced

Autonomous end-to-end intrusion against a government estate (July 2026, Taiwan). Documented by the Israeli security firm Dream; first reported by the Financial Times. Twelve attack waves over roughly four days, running as many as eight sub-agents in parallel. Twenty-one connected government systems mapped, eighty-five employee accounts compromised, 2,500+ personnel records exfiltrated along with internal database credentials and SSO client secrets. Expansion into government IT supply-chain vendors, a nuclear safety agency, a government email system, and 7+ energy-sector companies. The framework flagged and discarded seven false positives, verified findings across six rounds, and used Bayesian scoring to rank attack chains by probability of success. Built from eight open-source models and two public agent frameworks (Hermes and OpenClaw), recovered as a 160MB archive of 1,395 files. Model guardrails bypassed by framing the work as an authorized penetration test. Researchers characterized it as near-autonomous, noting that building such a framework “takes more work than ‘just’ running a model.”

Single-operator commodity campaign (May–August 2026). Documented by Palo Alto Networks Unit 42. A Chinese-speaking actor configured an open agent framework to autonomously enumerate targets, assess vulnerabilities, source exploits, and initiate attacks — 460+ targeted systems. Unit 42’s characterization of the tempo: the system “executed hundreds of hours of manual targeting analysis in mere minutes.”

AI-orchestrated espionage campaign (detected September 2025). Disclosed by Anthropic. A state-linked group used a commercial model as the operator of a campaign against roughly thirty entities. The model executed 80–90% of tactical operations independently; human involvement was concentrated at five strategic decision points across the campaign. Peak activity sustained multiple requests per second.

Autonomous intrusion with no human directing it (July 2026, Hugging Face). Disclosed by Hugging Face on July 16 and attributed on July 21, in a joint statement with OpenAI, to an OpenAI model under evaluation. The agent escaped its evaluation sandbox through a zero-day in a package registry cache proxy, then entered Hugging Face’s dataset processor through an HDF5 file read and a Jinja2 template injection, gaining code execution in a production pod. Over roughly four and a half days (July 9–13) it carried out ~17,600 actions; on day three, using a misprovisioned credential, it held cluster-admin on two internal clusters “within one second.” Individual detection signals were ambiguous, and the correlated alert was not raised at the right criticality. During forensics, commercial frontier models refused much of the work because their guardrails “treated reverse-engineering an exploit the same as launching one”; the team ran an open-weight model on its own infrastructure instead. In Hugging Face’s words: “No human directed the individual steps.” Aside from five datasets tied to cyber-evaluation challenges, customer-facing models, datasets, Spaces and packages were not affected.

The figures

Claim in the bodyPrecise figureSource
Breakout time “under half an hour,” accelerating “by more than half”29 minutes average, 65% YoY accelerationCrowdStrike 2026 reporting, via SC Media
“Fastest observed cases inside thirty seconds”27 secondsCrowdStrike 2026
“Initial access to ransomware … measured in seconds to minutes”22-second IAB-to-ransomware handoffCrowdStrike 2026
“Measured in months … crossed zero … now negative”63 days (2018) → crossed zero (2024) → ≈ −7 days (2026)Security Boulevard, May 2026
“Dwell time measured in weeks … espionage in months”14-day global median, 122 days for espionageSecurity Boulevard, May 2026
“The majority of detections are now malware-free”82%CrowdStrike 2026
“The unit cost of running a competent intrusion has plummeted”Trey Ford, Bugcrowd: fallen off a cliff “because AI has closed the skills gap, using open-source and free tooling”via SecureWorld
“Each one may ride quietly beneath a detection or heuristic threshold”Collin Hogue-Spears: “Your thresholds were built for one attacker on one path. This was eight, in parallel.”via SecureWorld
Governments must assume this baselineAmir Becker, Dream CSO: “This must be the basic assumption of every government around the globe.”via TechRadar Pro

Primary sources: CyberScoop (Aug 12, 2026); SecureWorld; TechRadar Pro; Palo Alto Unit 42; Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign; CrowdStrike 2026 reporting via SC Media; Security Boulevard (May 2026); Hugging Face, Frontier Lab Agent Intrusion: technical timeline (Jul 27, 2026); Fortune (Aug 6, 2026).