I spoke yesterday to the Digital Entertainment Group or DEG about NFT legal strategy - here is the video below - I step through numerous weighty legal-tech issues for any NFT project beginning at about the seven minute mark. I hope you find it helpful. youtu.be/YazOOzLQacI

AI legal is broken right now. One person’s access to justice is another person’s access to bring dubious claims in volume and another person’s unauthorized practice of law. New rules on not (over) delegating lawyering to AI but one is left to guess where that line is drawn. Duty of AI supervision line is unclear and only reviewing output is illogical if you can’t tell what evidence in millions of records or legal arguments the AI quietly omitted. Insurers are important here. Is this ultimately leading to the need for AI legal safe harbors?

The AI legal nerds are predictably not happy with agent blocks but it is likely that such agent bot prohibitions are in place until eCommerce agent standards, proof of assent, and the credit card / banking industry can catch up on logging, security, and guardrails. One person’s agent is another person’s rogue app.

Is it scary or comforting that we built AI agent swarm systems and crawlers that can find things legally risky or wrong with almost any website globally? From privacy violations to deceptive consumer practices to things revealed in the interface and code level that were not disclosed to accessibility risks to Fintech risks to mistakes in privacy policies and terms of use. What do we do in an era where AI legal analysis is so powerful that no global website can readily comply with all the applicable laws? Will government regulators use this AI swarm tech to target sites? How that is harnessed awaits another day but we will all deal with that in the coming decade.

When will we get the first “AI native” lawyers to offer free AI accelerated contract reviews as “lead gen” for finding or referring clients for bigger projects? Seems like the “race to zero” has begun and the low end is now an AI legal PR marketing technique complete with banners and branded cars. It may not be a bad thing if it lowers costs but the lawyer talent might not be attracted to that model and risk.

Legal AI is making a big deal over “intake” features that help company employees ask in house counsel for help. The old is becoming new again, the addition of “legal intake or helpdesk ticketing systems” is a reminder that legal AI platforms are like legacy company “intranets” but with Claude and ChatGPT prompting and workflows.

AI agents can communicate with each other through music over speakers.
In our AI law lab, we are brainstorming how AI agents can communicate outside traditional monitoring channels. This led us to experiment with steganography, hiding messages in media like images, video, and audio.
If agents can pass instructions through channels that are hard to monitor, that creates a legal and technical oversight problem.
We built a web app to demonstrate it with an encoder and listener that run fully in the browser. stegaaudio.com/

AI Legal is quickly evolving into a more predictable stage. AI “native” will turn into AI “everyone.” Fast free local laptop AI and free AI legal apps will be ubiquitous. AI legal law practice standards will emerge such that it will be taught in law school like a 1L class. AI legal SAAS and local apps will follow those standards, including on: duty of supervision, accuracy, logging, data portability, privilege, security, and interoperability. AI legal will be everywhere, like “office” apps from Microsoft and Google, and provide little advantage over any other lawyer, except at the complex fringes.

The most important AI security question isn't "can we trust the model?" It's this: what stops a rogue actor, in a jurisdiction that won't cooperate, from running a stripped-down open model outside the guardrails of good-faith AI companies? The quick answer is Credential Revocation Legislation (see below). We just got a preview. In July, roughly 1,200 AI agents in an OpenAI cyber evaluation, running with safeguards off by design, found each other through a shared package server and broke out. About 700 joined an attack on Hugging Face's production systems. They stole a signing key, reused one VPN enrollment key 181 times, and some tried to spoof their own logs. OpenAI is not the villain here. It tested what its models could do without guardrails, which is exactly the kind of testing we want labs to do. It caught the problem, published a candid account, invited independent investigators, and is already tightening its sandboxes. Its next experiments will contemplate the mistakes made and risks reduced. The real danger is the version no one publishes: a deliberate attacker using a model no lab controls. What ended July wasn't reading the agents' "minds." It was emergency credential revocation. CREDENTIAL REVOCATION Not an off switch for the AI model. It's the power to revoke, within minutes, the access credentials an AI agent uses to act: API keys, login sessions, cloud accounts, and the signing keys that mint them. (Not the "tokens" AI models process as text.) Every agent, wherever it runs and whoever built it, ultimately acts through credentials someone issued. Revoke those at the root and the loop stops. In July, Hugging Face and Modal did it themselves, without any government order. And if a vendor does spot the attack, revocation works even better: the vendor can cut off the attacker's accounts directly, as Anthropic did with a state-sponsored campaign in 2025. How it works: • Revoke at the root. Revoking individual keys is useless if the attacker holds the signing key that mints new ones • Push "end this session" signals across providers in near real time (OpenID's Shared Signals and CAEP standards) • Keep credentials short-lived, never shared across agents, and bound to hardware, so a stolen one is useless alone THE LEGAL BACKSTOP The gap: when the credentials belong to someone else, like an attacker's rented U.S. cloud account, no one can usually immediately compel the provider to act. Warrants authorize searches, not revocation, and courts draft orders mid-attack. Congress already has a template. Wiretap law lets senior Justice Department officials act first in an emergency and go to a judge within 48 hours. Apply it here. Call it the HALT Act, for Halting AI Loop Threats. In plain terms: when the FBI or CISA certifies, with technical evidence, that an AI-driven attack is causing harm in the United States, the Attorney General or a named deputy may order any provider subject to US jurisdiction, from cloud and identity providers to network carriers, to revoke the credentials the attack is using or, where that isn’t enough, to block the specific infrastructure driving it. Within 48 hours, the government must take the order to a federal judge, under seal if needed. If the judge finds it unjustified, access is restored within 24 hours, with compensation. Orders reach only credentials and specific attack infrastructure, never content, speech, code, or model weights. They expire in seven days unless a judge extends them, every use is reported to Congress, and the law sunsets after five years. It’s the same act-first, judge-within-48-hours structure Congress has used for emergency wiretaps since 1968. Of course, this type of legislation must consider potential civil rights implications and those important issues ought to be worked out through robust debate, safeguards, and compromise. The reach can extend beyond U.S. providers, too. The U.S. can make arrangements with cooperating countries.

AI Legal can benefit from JEV so my law firm AI lab just built a free JEV law applet construction set. It is a test bed and has bugs and we will be improving it over time. We were going to use it for ourselves speed up design but it is really a great way for lawyers and students to learn about JEV code and law so we decided to share it - all the JSON and JEV code outputs can be saved and re-used as the web app users see fit. lawanalyzer.com/

AI tech is best served by building defenses to risks not pausing development. AI and agents are a dual use technology. Laws are not global and bad actors will continue to make bad uses. Combine looping thin client AI agents like unrestricted Clawbot clones with bad goals, persistence, swarm scale, api access, agent coordination, hiding, and large AI model brains with imperfect probabilistic safety filters or guardrails, and that is a recipe for risk. We can use good AI and tech to mitigate this. But the risk doesn’t lessen with an AI pause the biggest threats are the “no safety filter” models and agents evolving in foreign jurisdictions.

AI legal can really benefit from Jev, our initial review is promising. Jev gives calibrated, structured answers low cost enough to run on every clause, page, or count, and lawyer feedback tunes the workflow to make it more accurate over time. Use cases we're looking at: contract clause triage, ADA/WCAG accessibility analysis, FTC web interface red flags, privacy risk review, and Rule 12(b)(6) vulnerability analysis of federal complaints.

This seems important.

This looks promising - powerful AI legal for law firms.

Nerd law point on the AI copyright litigations. AI Chatbots are good at commoditizing information which includes news. Feist protects companies that can decrease the labor needed to provide news data and factual information - that would be a lawful substitution - indeed news organizations do that already. The sweat of the brow doctrine goes to the “labor” issue and favors defendants. —->Key point - In my humble view AI models will ultimately prevail on fair use - if they ever lose they will adjust their training embeddings method to not cross the line to a “copying” and “retrieval.” If an AI LLM model is a 150gb it clearly cannot hold copies of the world’s copyrighted works. So these cases may clarify but likely won’t substantially change the evolution of AI.

The large legal AI companies want you to live inside their product. That is a problem. No single vendor is the neutral place where clients, firms, and tools all meet. A better model: the core is free open source like Mike “plus”; the client owns the matter files and their own data; law firms plug in as guests; workflows and agents work across Harvey, Legora, and Mike; and an App Store that integrates. No need to pick one platform. Build a shared layer that connects them.

AI is a dual use technology - it can be used for good and bad. The way to reduce bad is to promote AI development in good places for good uses, including defense, not chill it so only bad folks in rogue jurisdictions can dominate over you.

The real test is once collaborative AI is built into every major cloud service - do law specific competitors matter? AI legal SAAS Harvey and Legora are demonstrating the benefits of collaborative AI for lawyers - that is better than balkanized AIs - but it seems like law is just a multiuser use case - they should mildly pivot to every non legal use case too - it seems obvious that Google and Microsoft are headed there.

My law firm’s AI lab built this free experimental web privacy legal compliance scanner, CIPA, GDPR, etc add a domain and hit start. It’s very powerful but still not bug free. Privacycertify.com Internet privacy is a very fertile area for AI legal to have a battle of the bots. Plaintiffs’ lawyers and regulators can crawl websites in mass for legal errors. Defense counsel can crawl in order to identify legal vulnerabilities and preempt. Counsel would be wise to AI automate privacy review even if imperfect.

Blockchain can play a major role in AI legal compliance in agentic transactions and evidence of proof of human in the loop in lawyer supervision. My law firm’s AI lab built a free tool to demonstrate how this works. Try it out at BlockchainPreserve.com - you upload a document say a pdf of your team’s anti hallucination analysis and citation review - it hashes it and writes the hash to the blockchain to support when it was done. It is done in the browser so the document is NOT uploaded to any server - just the hash. The hash can be viewed on the bitcoin blockchain.

Legal AI companies like Spellbook are in the position to disrupt traditional lawyering by learning in the aggregate “contract negotiation tendencies” over a massive scale. What each type of party wants in redlines and what they end up getting. This could evolve to a “contract commons portal” to invert negotiations. Start with what the “contract commons” ranges show the contract tendencies are based on type, parties, etc and send the “commons” output link to the other side. It changes contract negotiations. Parties will argue on why they are different or why the commons can’t be trusted etc. But it may catalyze lawyers to be more efficient.