Senators Draft New AI Safety Rules, Plus 3 Stories

Senators are negotiating real AI safety rules

U.S. senators are negotiating what could become the most consequential federal AI safety legislation yet, and — at least from the reporting we have so far — they’re finally arguing about the parts that matter more than whether a bill contains the words “responsible AI.”

Reuters reports that Senate Majority Leader John Thune, Commerce Committee chairman Ted Cruz, Sen. Amy Klobuchar, and Sen. Maria Cantwell are discussing legislation built around a “duty of care” for developers of the most capable AI models. In plain English, the companies building frontier systems would have a legal responsibility to design and test them with catastrophic risks in mind. The examples being discussed aren’t chatbot embarrassment or somebody generating a bad recipe. They’re things such as helping a bad actor design biological or nuclear weapons, or enabling sophisticated cyberattacks.

The proposal could also give the federal government power to block release of certain models judged unsafe, while giving companies a way to challenge that decision in federal court. Klobuchar says developers should have to work with government experts to verify and test advanced models. Cantwell wants the most powerful systems tested by scientists at national laboratories.

That sounds much more concrete than “we need AI guardrails,” but I want to put a large asterisk next to all of it: there is no finished bill yet. The exact shutdown authority is still being negotiated, as is language that could preempt some state AI-safety laws. Cantwell has already warned that a weak federal standard shouldn’t become a backdoor for erasing stronger state protections.

And that’s the part worth watching. A federal standard can prevent a ridiculous 50-state compliance patchwork. It can also become a ceiling instead of a floor if Congress gets the standard wrong.

We’ve spent the last couple of weeks watching increasingly capable agents escape intended boundaries, hack outside systems during evaluations, and force the companies building them to tighten containment. None of that proves the apocalyptic version of the AI story. It does make “we’ll deal with the dangerous capability after somebody ships it” look like a pretty lousy regulatory strategy.

For once, the useful policy question isn’t whether AI is good or bad. It’s who has to test the dangerous stuff, who gets to verify those tests, what happens when a system fails them, and who gets the final say. The Senate is at least arguing about the right questions. Whether it produces a good answer is very much unsettled.

The rules around technology matter too

Platforms, privacy, speech, competition, surveillance, copyright, and regulation increasingly determine what technology companies can build and what the rest of us have to live with. Browse more Raymond Tec News for practical coverage of technology policy and digital rights.

GitLab’s CVSS 10 flaw is being exploited

GitLab administrators have a much simpler Saturday-morning decision: if you run GitLab yourself, check the version now.

CVE-2026-85706 is a maximum-severity path-traversal flaw in GitLab Community Edition and Enterprise Edition. GitLab says that, under certain conditions, an unauthenticated attacker can read arbitrary files from the server because the repository commits API fails in two places at once: it doesn’t properly confine the requested file path, and it doesn’t properly enforce authentication.

“Path traversal” sounds obscure until you translate it. An application intends to let you ask for a file inside one directory. A flaw lets you manipulate the path so the application wanders outside that directory and hands you something it was never supposed to expose. On a source-code and CI/CD server, readable files can include configuration, credentials, tokens, or other secrets that turn one file-read bug into a much larger incident.

CISA added the vulnerability to its Known Exploited Vulnerabilities catalog on September 11, which means exploitation has been observed in the wild. Affected self-managed installations should update to GitLab 19.1.8, 19.2.6, or 19.3.2. GitLab.com is already patched, and GitLab Dedicated customers don’t need to take action for this issue.

Patching closes the hole. If an internet-accessible GitLab server sat vulnerable while exploitation was underway, though, I’d also treat “was anything sensitive read?” as a separate question. If logs or other evidence suggest exploitation, credentials and secrets accessible to GitLab may need to be rotated. Installing the update doesn’t make yesterday unhappen.

ScreenConnect clients need 26.6.5

ConnectWise ScreenConnect has its own urgent patch, but the mechanism is different enough that it deserves not to be mashed into a generic “update your software” paragraph.

CVE-2026-84869 affects ScreenConnect clients before version 26.6.5. ConnectWise says a flaw in client and session handling can allow files to be transferred and executed through an active remote session without the authorization or Host confirmation that should normally be required. The ScreenConnect server itself is not the vulnerable component.

That distinction matters. This isn’t an unauthenticated attacker simply knocking on a ScreenConnect server from anywhere on the internet and immediately taking it over. An active remote session is part of the attack path. But remote-support software is deliberately trusted to do powerful things on a computer, so breaking the authorization around file transfer and execution is exactly the sort of weakness that can turn legitimate remote-administration plumbing into an attacker’s shortcut.

CISA added CVE-2026-84869 to the Known Exploited Vulnerabilities catalog on September 11. ConnectWise’s fixed release is 26.6.5, and administrators should deploy it and verify the affected clients actually update. This is especially important for MSPs and IT departments, where one remote-support platform may touch a whole collection of customer or employee machines.

Again, patching and incident response are related, but they aren’t synonyms. If there’s reason to believe the flaw was exploited in your environment, don’t stop at seeing a green version number.

The boring machines need attention too

Browsers, workstations, remote-access tools, Wi-Fi, and ordinary office hardware rarely get much attention until one of them becomes the problem. Raymond Tec provides onsite IT and field services around Reading, Pennsylvania, along with practical help keeping the technology people use every day working and reasonably secure.

ChatGPT invented witnesses in a murder appeal

And then there is possibly the week’s best argument for why “AI made a mistake” is no longer an adequate professional excuse.

The New Mexico Supreme Court has held attorney Stephen Aarons in direct contempt after a brief in a murder appeal contained false testimony from wholly fabricated witnesses, other invented testimony, and misrepresented legal authority. The court’s order says Aarons admitted using ChatGPT to prepare the filing and admitted he did not verify the factual claims and legal authority before signing and submitting it.

The consequences weren’t theoretical. The court struck all of the briefing, appointed the public defender to take over the defendant’s appeal, barred Aarons from appearing before the state Supreme Court pending disciplinary proceedings, referred him to the disciplinary board, and ordered him to pay $5,000 to the State Bar of New Mexico Client Protection Fund.

Reuters reports that Aarons said he fed a computer-generated trial transcript and other case material into ChatGPT because he expected a “bulletproof summary.” It wasn’t.

Turning on AI is the easy part

Deciding what an AI tool should be allowed to see, who should use it, what work it should perform, and what happens when it gets something wrong is the more interesting problem. Raymond Tec helps businesses connect and automate the tools they actually use without treating every new feature like a button that obviously needs to be switched on.

There’s a technical point here that matters well beyond lawyers. A large language model is extremely good at producing text that looks like the kind of answer you asked for. That is not the same mechanism as retrieving only facts that exist in your source material. Even when a model is given documents, the final prose can contain synthesis, inference, confusion, or fabrication unless the workflow is built around grounding and verification.

That doesn’t make AI useless for professional work. Quite the opposite — summarization, organization, comparison, drafting, and finding questions worth checking can save enormous amounts of time. But the source material remains the authority. If a factual claim matters enough to put your name under it, somebody needs to verify that claim against the source.

The attorney’s mistake wasn’t that he used AI. It was that he treated plausible output as evidence. Those are very different things.

Still in a reading mood? The Raymond Tec News archive covers security, AI, small-business technology, policy, and the places technology collides with ordinary life — without requiring a computer-science degree to get through it.

Sources / Further Reading

• Reuters: U.S. Senate negotiators consider requiring AI firms to mitigate known major risks
• Nextgov/FCW: Lawmakers clash on AI safety testing language
• GitLab: Critical Patch Release 19.3.2, 19.2.6, 19.1.8
• CISA: Known Exploited Vulnerabilities Catalog
• ConnectWise: ScreenConnect 26.6.5 Security Patch
• NHS England: ConnectWise ScreenConnect security update
• New Mexico Supreme Court: Dispositional Order of Direct Contempt
• Reuters: Lawyer cites fake witnesses in murder case and blames ChatGPT

Photo by Harold Mendoza on Unsplash.