NSA, CISA, and four allied agencies just put something in writing that usually stays classified: the AI models capable of breaching your firm's defenses are months away, not years. They didn't say this to other spy agencies. They said it to boards and CISOs — the people who set budgets, not the people who read logs.
THE LEAD PLAY
Your Government Just Told You How Much Time You Have Left
On June 22, the cybersecurity arms of the US, UK, Canada, Australia, and New Zealand — NSA, CISA, GCHQ, and their counterparts — signed a joint statement most of the security world is still catching up on. The headline line: frontier AI models are expected to fundamentally transform offensive and defensive cyber capability, and the timeline for that is months, not years.
That's not vendor marketing. That's six named officials, including NSA's cybersecurity director and CISA's acting director, putting their names on a document that says the assumptions most security programs run on — how fast a vulnerability turns into an exploit, how much warning you get — are already out of date.
The model the agencies are describing sounds like the one that made your radar three weeks ago. CyberScoop reported the warning ties directly to the kind of capability found in Anthropic's Fable 5 — the same model the government pulled from the market in Issue #2, then let back online after Anthropic agreed to new safeguards. Same underlying capability. Different agency, different memo, same conclusion: this isn't theoretical anymore.
Law firms and legal departments hold exactly the material this warning is about — privileged communications, deal terms, litigation strategy, HR investigations. You're not a bystander to this. You're inventory.
The Play this week: Get 30 minutes with whoever owns your firm's or department's incident response plan and ask one question: does it assume an attacker with AI-accelerated capability, or does it assume the threat landscape from two years ago? If nobody can answer that clearly, that's the gap. You don't need to solve it this week. You need it on the agenda before "months" turns into "now."
If you outsource security to a vendor — MSP, cloud provider, practice management platform — ask them the same question you'd ask about an AI model vendor: what changes about your posture if the threat timeline the government just published is right? Get the answer in writing. You're going to want it the next time a client's security questionnaire asks.
SUPPORTING PLAY 1
Texas and Connecticut Are Writing the Rule You Already Guessed At
Issue #1 covered Florida's Rule 2.515 — signers certifying that cited authorities exist. Texas's Supreme Court has now proposed its own rule changes addressing AI misuse, including sanctions and a signature-attestation requirement of its own. Connecticut's attorney rules committee is reportedly moving on similar ground: sanctions serious enough to end a case over AI citation errors.
Three jurisdictions, three separate rulemaking processes, same direction. This stopped being a Florida quirk. It's becoming the baseline.
If you took Issue #1's advice, you already have a written, tool-agnostic verification protocol. Good. Don't file it. Pull it back out and check it against what these rules actually require — not what you assumed a reasonable protocol should cover. Texas and Connecticut are going to specify exactly what a signer has to certify, in exact language. A protocol that doesn't produce evidence mapping to that language is one that makes you feel better without giving you anything to point to in court.
The Play this week: Pull your citation-verification protocol. Check whether it produces a record — not a feeling — that a signer could point to if a judge asked "how do you know." If it doesn't, that's the fix, not a rewrite.
SUPPORTING PLAY 2
Kirkland Just Showed You What Losing Looks Like
Kirkland & Ellis signed a multiyear deal with Syllo — the litigation AI platform already used by firms like Ballard Spahr, Pillsbury, and Quinn Emanuel — that gives Kirkland exclusive rights to build custom tools and knowledge infrastructure on top of Syllo's platform for an extended period. Syllo isn't leaving the market; other firms can still use the core product. What Kirkland locked up is the layer that turns a shared platform into firm-specific infrastructure.
This is the same firm that committed $500 million to its own AI build-out and partnered with Palantir on fund formation. None of that is news to you at this point. What's new is the shape of this deal: exclusivity on the build layer, not just adoption of the product.
Nobody at a mid-size firm is competing with Kirkland for Syllo's attention. That's not the point. The point is that a vendor worth having was willing to sign away its most valuable customization rights to one client. The base product might stay open — the ability to build something differentiated on it might not.
The Play this week: If you're seriously evaluating a litigation AI vendor, ask directly whether they've discussed exclusivity or extended custom-build rights with anyone else. You're not trying to out-bid Kirkland. You're trying to find out whether that option is still on the table before someone else takes it.
QUICK HITS
A Connecticut federal judge told attorneys to push back on their own clients. At a sanctions hearing, the judge said generative AI is no substitute for professional judgment — it tends toward telling you what you want to hear and can get the law wrong — and specifically urged lawyers to resist clients who demand AI-driven research. Worth having an answer ready the next time a client asks why you're not just running something through AI yourself.
The export-control fire drill from Issue #2 is turning into a standing process. The White House is reportedly close to a voluntary framework with OpenAI, Anthropic, and Google that sets fixed benchmarks and review timelines for frontier model releases, with an announcement possibly landing as soon as this week. If it does, the vendor-dependency mapping from three weeks ago stops being a one-time fire drill and becomes something you check on a schedule.
Litera published a survey of roughly 100 general counsel at Fortune 1000 companies arguing AI governance hasn't kept pace with AI adoption. It's vendor content marketing more than research — but the underlying question is real. If a board member asked what your AI governance actually looks like, do you have a specific answer or a vibe?
That's Issue #3. Find out who owns your incident response plan, then ask them when it was last opened. See you in the next one.