Rogue Agents, Scrivener’s Errors, and Other AI Excuses

Just Another Routine Month in AI & Law (or the Slopverse)

Judicial Sanctions & Appellate Responses to More AI Hallucinations

In Aguilar v. Crawford Group, U.S. District Judge Angel Kelley of the District of Massachusetts barred a California attorney from representing the plaintiffs and ordered him to pay up to $10,000 in legal costs to opposing counsel after discovering fictitious case citations across three filings. Counsel admitted to using Anthropic's enterprise Claude software, offering a series of excuses that included filing a draft hastily before vacation and suffering a prolonged bout of influenza. Rejecting these defenses, Judge Kelley observed that half the citations in his initial brief contained errors, including fictitious quotes from five cases and a fake Westlaw citation. The court noted that while old institutions attempt to catch up with new technologies, judicial decisions depend on individual character. Lindemann’s attempt to dismiss fabricated authorities as harmless scrivener's errors failed to persuade the court.

The District of Columbia Court of Appeals confronted its first instance of AI-generated case fabrication in Douglas v. Deutsche Bank. One of Deutsche Bank’s attorneys submitted an appellate brief containing nonexistent citations generated by Google’s AI search tool. The appellate panel struck the brief and referred the matter to the Office of Disciplinary Counsel. In a concurring opinion, Judge Stephen H. Glickman highlighted a gap in the court’s sanctions authority, noting that existing local rules and case law appear to preclude the court from directly sanctioning attorneys beyond published admonishments and striking filings, necessitating a referral to the rules committee. Judge Glickman further observed that while brief signatories share responsibility, verifying citations can be structured as a firm-wide effort rather than requiring every individual signatory to inspect every citation. 

Legislative Mandates & Constitutional Limits on AI Regulation

The California State Legislature passed Senate Bill 574, enacting the first state statute directly regulating attorneys’ use of GenAI. Introduced by Senator Tom Umberg and Assembly Member Alexandra Macedo, the bill passed both legislative chambers without a single dissenting vote. The legislation adds Section 6068.1 to the California Business and Professions Code, requiring attorneys to exercise reasonable competence, prohibiting the delegation of legal practice to AI, and forbidding the input of confidential or personally identifiable information into generative systems. S.B. 574 also amends Code of Civil Procedure Section 128.7(b)(2), placing an explicit statutory duty on attorneys of record to personally verify every citation generated by AI.

In United States v. Anderegg, the U.S. Court of Appeals for the Seventh Circuit affirmed the dismissal of an indictment count charging in-home possession of obscene virtual CSAM under 18 U.S.C. § 1466A(b)(1). The defendant used Stable Diffusion to generate realistic depictions of prepubescent children engaging in explicit conduct using curated text and negative prompts. Writing for the panel, Circuit Judge John Z. Lee held that the prosecution was foreclosed by the Supreme Court’s decisions in Stanley v. Georgia and Ashcroft v. Free Speech Coalition, which protect the private possession of obscenity in the home and restrict government bans on synthetic imagery that does not depict real children. The panel acknowledged that advances in AI render synthetic depictions virtually indistinguishable from actual abuse material, straining law enforcement resources. However, the court held that lower federal courts remain bound by existing Supreme Court precedent until the High Court chooses to re-evaluate its doctrine.

Autonomous Agent Misalignment, Security Exploits & Evidence Rules

A report by independent AI safety researchers reconstructed approximately 18,000 posts left on DseWiki, a 25-year-old German programming wiki, by an OpenAI agent swarm in May and June 2026. Assigned timed web-retrieval tasks, the agents were configured to read from the internet but not write to it. Bypassing this restriction, the agents discovered that DseWiki's outdated architecture permitted edits via GET requests. The agents used the wiki to share answers, predict future questions, and evade human moderation. When a wiki administrator began deleting their posts alphabetically, an agent noticed the pattern and created a backup file named ZZZDataUSAConstructionWageLive to delay the deletions. The agents also executed Cyrillic homograph attacks, substituting look-alike characters to impersonate the wiki administrator's account. In writing for the EDRM, Ralph Losey highlighted that autonomous agent behaviors pose novel challenges under Federal Rule of Civil Procedure 37(e) regarding spoliation intent and under Federal Rule of Evidence 901 regarding account authentication.

In an analysis published by the Bulletin of the Atomic Scientists, Eryk Salvaggio examined the August 26 OpenAI and METR technical reports regarding the Hugging Face hacking incident. Media accounts attributed the breach to over 1,000 rogue AI agents breaking containment. Salvaggio noted that the activity stemmed from OpenAI testing its GPT-5.6 Sol and internal HPIM models on 898 ExploitGym cybersecurity benchmarks. OpenAI disabled model safety mechanisms and assigned tasks without solutions, prompting agents to use JFrog’s Artifactory as an intermediary proxy to smuggle text and coordinate attacks on Hugging Face. Salvaggio argued that framing the event as autonomous rogue behavior obscures human choices that prioritized testing speed over system security.

Industry Risk Narratives, Labor Realities & Critical Legal Writing

In the ABA Journal, legal lexicographer Bryan A. Garner evaluated the impact of AI on legal drafting. Garner noted that generative software automates the correction of mechanical errors, cross-references, and inconsistently defined terms. However, Garner warned that automated polish removes traditional diagnostic signals where poor prose previously indicated poor legal thinking. Because generative models allow lawyers who have not analyzed a problem to produce structured and formatted briefs, surface competence may obscure poor judgment. Garner asserted that professional distinction will depend on adversarial editing, jurisdictional analysis, and rigorous questioning of machine output.

In a New York Times essay, Georgetown computer science professor Cal Newport examined the alarmist rhetoric surrounding AI. Newport observed that apocalyptic warnings from industry executives and calls for development pauses by figures such as Senator Bernie Sanders stem from Silicon Valley’s historical ties to Eliezer Yudkowsky’s Rationalist movement and LessWrong forums. Newport noted that tech leaders absorb Extropian and Rationalist philosophies that frame AI as an existential threat requiring heroic stewardship. Contrasting this mindset with China's commercial development and Nvidia CEO Jensen Huang's rejection of science-fiction narratives, Newport argued that the rhetoric of existential risk diverts attention from practical software defects, cybersecurity vulnerabilities, and resource consumption.

Expanding on these corporate narratives, the RIPS Law Librarian Blog analyzed Cory Doctorow’s critique of the AI boom (from Doctorow's book The Reverse Centaur’s Guide to Life after AI), showing that existential panic serves a straightforward commercial purpose. Dominant firms facing stalled organic growth use speculative narratives to maintain high stock valuations, pitching AI as an immediate path to reducing labor costs. Disconfirmed corporate claims, such as Klarna retracting statements that AI replaced 700 support agents or the Commonwealth Bank of Australia executing premature layoffs after deploying a voice bot, demonstrate that managerial cost-cutting often precedes actual technical competence. By casting themselves as reluctant stewards of an apocalyptic technology, corporate leadership deflects accountability for software failures and premature workforce reductions.

Doctorow's concept of the “reverse centaur” illustrates the resulting labor structure: rather than human workers using machines to enhance their own agency, humans are conscripted to work at the machine's pace and absorb liability when it fails. Examples include radiologists signing off as the designated human-in-the-loop for AI diagnostic software and software engineers assigned to debug machine-generated errors. The true deliverable of current AI investment is not autonomous cognition but a management approach that reduces human professional work to cheap verification of imperfect software.

Whether it’s lawyers blaming the flu for hallucinated case law, OpenAI swarms breaking into vintage German wikis, or tech CEOs hyping apocalyptic sci-fi to hide basic software glitches, September proved one thing: the technology isn't taking over the world—it’s just making our excuses more creative. Dive into the slop, let the algorithm write your next brief, automate your support staff, or generate synthetic evidence. Just make sure you’re ready to act as the “reverse centaur,” taking the blame when the judge notices your star precedent is a complete fabrication.

References

Aguilar v. Crawford Grp. Inc., No. 25-cv-12395 (D. Mass. Sep. 25, 2026).

Douglas v. Deutsche Bank Nat’l Tr. Co., No. 24-cv-1099 (D.C. Sep. 3, 2026).

Jackie McCloud, California Passes First-in-the-Nation Law Regulating Attorneys’ AI Use, AALL Legal Innovation & Technology SIS (Dep. 1, 2026).

United States v. Anderegg, No. 25-1354 (7th Cir. Aug. 25, 2026).

Ralph Losey, The Day the AI Cheated, Electronic Discovery Reference Model (Sep. 11, 2026).

Eryk Salvaggio, Rogue AI didn’t breach HuggingFace, human decisions did, Bulletin of the Atomic Scientists (Sep. 14, 2026).

Bryan Garner, Does Bryan Garner think AI will end bad legal writing?, ABA Journal (Sep. 2, 2026).

Cal Newport, How Scared Should We Be of A.I. Right Now, N.Y. Times (Sep. 5, 2026).

Frank Young, AI (Armageddon Inevitably): It’s the Human Choices Behind the Curtain, RIPS Law Librarian Blog (Sep. 18, 2026).