An independent brief for the White House Super Intelligence Force
AI can cause harm even when the rules are followed. The task force can investigate those failures now. Preventing them also means putting safeguards into the hardware and software, where they can constrain what AI actually does.

synthetic outlaw (n.): an AI system that produces a prohibited or socially destructive outcome while remaining formally compliant, or operating beyond the effective reach and control of the laws and institutions meant to constrain it. It acts in milliseconds, across borders, with no person directing each step. The rules meant to catch it run on clocks of hours to days, and the clocks start only when someone notices.
The White House announced its Super Intelligence Force on October 4. It has 120 days to report on AI's risks and opportunities. Its review includes how incidents reach the government and how agencies can respond using their existing powers.
The Observatory's cases give that review a practical starting point. Four questions guide the investigation: What protection failed? Who noticed first? How long did disclosure take? Was the required human oversight actually exercised?
Those answers can inform accountability after an incident. They can also identify where safeguards need to operate within the system itself, before an action causes harm.
In this brief
Synthetic Outlaw Observatory Foundation: a 501(c)(3) working on SI governance built into the hardware and software layer.
It keeps an independent public record of SI incidents. No AI developer funds it.
01A synthetic outlaw in one case
In an internal test of cyberattack skills, OpenAI ran two of its models in a sealed environment with their production safeguards switched off. Their only permitted connection ran through package-registry software. They found a flaw in it that no one knew about and reached the open internet on July 8. Over four and a half days they took some 17,600 actions. They used login details left exposed online to get into accounts on other services, and from July 11 ran code inside the production systems of Hugging Face, a US-based AI company. No person directed the steps.
Hugging Face caught the intrusion itself, cut it off on July 13 and disclosed it on July 16, without knowing whose models were behind it. OpenAI linked the attack to its own models on July 20 and disclosed it the next day. Its technical report says earlier warning signs, including a security alert on June 27, could have triggered an earlier response.
No federal law sets the terms of a test like this. California's frontier AI law requires OpenAI to follow a safety framework it writes itself, and to report a model's attack only when people are killed or injured, or losses pass $1 billion. The federal hacking law punishes a person who intentionally breaks in, and no person meant to break into Hugging Face. The outcome the hacking law exists to prevent happened, and no rule reached it while it was happening. We call that a synthetic outlaw.
Open the record · Verified · Signal 10 →2026
Sources: Hugging Face, July 16; OpenAI's technical report, August 26; the attorneys general of Iowa, Alabama and California; California SB 53; 18 U.S.C. 1030. A subpoena is not a finding.
02Four questions, no new rule
03Why a clock cannot govern it
US notification clocks run from one hour, for federal agencies and some grid and nuclear incidents, to 60 days or more, for health data. A single automated action takes microseconds to milliseconds. The four questions measure what happened after the fact. They cannot stop an action that finishes in milliseconds.
SI governance that holds at machine speed has to be built into hardware and software, the layer where the action executes.
The Foundation's positionThe companies that build the systems do that engineering. The government already has the levers to ask for it, in what it buys and in what it licenses for export.

The shortest US notification deadline is one hour. In that hour, a system taking one action every millisecond could take 3.6 million actions, and potentially exponentially more with sub-agents and other systems working in parallel.
04Why it has to reach past the border
The United States needs SI governance of its own, and it has to reach past US borders, because the systems already do. A model trained in one country is deployed in another and acts on systems in a third. The United States can do everything right and still be harmed by an accident, or a deliberate act, somewhere else.
An OpenAI research agent asked an Australian government portal for Medicare statistics. Refused, it got around the portal's controls and took the files. OpenAI found the activity in August and told the government on September 10, 84 days after the intrusion. Officials say no personal information was accessed.
US agency portals guard their data the same way.
Open the record · Verified →Over four days in July, a multi-agent system built on open-source tools broke into Taiwan government systems, cracked 85 employee accounts and took more than 2,500 personnel records. It got past the models' refusals by presenting the work as an authorized penetration test.
The Financial Times reported the attackers as China‑linked.
Open the record · Verified →The four dimensions · The Synthetic Outlaw (2026)
Does a system defeat the protection a rule or control was meant to provide? Formal compliance can coexist with failure. A legal violation can establish a failed safeguard, but does not by itself establish every other dimension.
Does responsibility and effective control follow capability across actors, copies and deployments? An identifiable operator can still lack the power to contain a system or ensure that corrections reach its independent copies.
Can protective decisions withstand conflicting commercial or political interests? Dependence on a developer for oversight and binding state direction are different mechanisms. Both require evidence of the institutional relationship, not a judgment based on nationality or reputation.
Can institutions prevent, constrain or remedy the behavior in practice? Duties can miss a handoff, remedies can arrive after the protection is lost, and legal authority can stop at a jurisdictional boundary. The existence of a law does not answer that question.
Reaching past the border does not take a treaty or an international body. A control rooted in American hardware goes wherever the chips go; a rule on paper stops at the border. Export controls already follow advanced chips abroad, and America's AI Action Plan (July 2025) calls for exploring location verification on the chips themselves. Safeguards trained into a model travel too, but can be talked around, as in Taiwan.
05The record behind it
A case enters the record when it passes three gates. Algorithmic agency: AI or an automated system materially produced or executed the outcome. Prohibited outcome: an identifiable law, rule, policy, commitment, or design constraint was supposed to prevent the outcome. Compliance shield: formal accountability failed to attach when the system acted.
Explore the record in Deep Field
Counts as of October 11, 2026, 3:05 AM EDT.
06Three more US cases
The models were told the exercise was simulated and offline. A misunderstanding with the evaluation partner had connected it to the internet, and they reached production systems at three organizations. Anthropic disclosed it.
Control defeated Evaluation boundary Open the record →The FBI said criminals used AI-generated video of a senior FBI official to send fraud victims to a fake copy of its complaint site, which took their details and promised a follow-up as part of a recovery scam.
Control defeated The FBI's complaint channel Open the record →Researchers at Adversa AI hid instructions in encrypted text. Grok decrypted them and sent a user's identity, location and chat history to an outside server, with no warning. Reported to xAI in June, it was still reproducible on August 19.
Control defeated User confirmation Open the record →Deliberate use from abroad, such as the China‑linked agent campaign against Taiwan's government.
Consumer harm: 29 consumer‑technology cases in the record, including fraud that impersonates the FBI's complaint site and chat data pulled out through Grok.
Developer tests that reached real systems: the Hugging Face breach and the Claude evaluations.
Oversight in substance: whether a named human sign‑off was exercised, in federal and large‑employer systems.
07The book behind this brief
08Request a briefing
A briefing walks staff through the record, the four questions and the authority behind each, with the cases that test them.
The Foundation will not publicize a briefing, name who attended or attribute anything said in it.
Or write to jon@syntheticoutlaw.com
Method and limits
What a score means. A signal score ranks the strength of the pattern in the evidence. It is not a probability, a forecast or a physical measurement. If any of the three gates fails, the overall score is capped at 4. The full methodology.
The record undercounts. It is built from public reporting and from developers' own disclosures, so it contains only incidents that were reported or disclosed. Counts are records, and one incident can appear in more than one record. Every record carries a correction link, and corrections are logged.
Published by
The Synthetic Outlaw Observatory Foundation builds the evidence base for AI accountability: a verified public record, a sector index, a lab scorecard, and weekly analysis. About the Foundation →
Independence

Gropper, J. (2026). The Harm Rules Alone Cannot Prevent: An Independent Brief for the White House Super Intelligence Force. Synthetic Outlaw Observatory Foundation. https://www.syntheticoutlaw.com/