> daily_signal(2026_09_02)
A cop turned a camera network into a stalking tool, an AI agent now reads your logins on install, Isbell sued Suno over his voice, and an AI designed a real mission to the nearest star.
PickBits Daily Signal · Wednesday, September 2, 2026
// tl;dr
- A Georgia police officer used Flock's nationwide license-plate-reader network to track a fellow officer he'd had an affair with, and a man whose car kept appearing near hers. Wired reported the abuse from internal investigation records on August 27. He didn't break into anything; Flock's cameras log every passing plate into a database any signed-in officer can search, and no per-query check on purpose stopped him.
- OpenClaw 2.0, the open-source AI-agent platform, now auto-detects your existing ChatGPT and Claude logins on install and wires itself in. The Decoder reported the release, over 16,000 pull requests, on August 31. Its new "cloud sessions" can run an agent off the laptop on a rented AWS or Hetzner box, carrying prompts, files, and any harvested credential with it.
- Jason Isbell filed a proposed class action against the AI music app Suno, accusing it of taking artists' musical likeness without consent and likening it to Star Trek's "Borg." The Hollywood Reporter reported the September 1 filing. Led by working musicians rather than record labels, the suit argues consent, not just copyright, is the line Suno crossed.
- An AI called "Get Physics Done" designed the flight path for a real, funded mission to Alpha Centauri, and a nonprofit committed about $15 million to launch it by the end of 2029. MIT Technology Review reported the plan on September 1. After roughly three autonomous days, the AI proposed swinging the craft closer to the sun than Mercury to gain thrust and cut its mass.
The Flock story is the one we've been circling for a month. Back on August 8 we covered the company's pitch to turn 350,000 Uber and Lyft cars into a rolling plate-reader network, and the question we left open was who checks the officer running a search. This week Wired answered it: a Georgia cop used Flock to stalk an ex, and nobody checked anything. The AI-music consent fight we've followed since June's performer rulebook got its sharpest plaintiff yet this week too, in Jason Isbell.
OpenClaw's new release is the one that'll land on your own desk. It makes an AI agent that grabs your logins on install feel like pure convenience, right up until it's a corporate credential leaving your building. And then there's the good one, which I'll admit surprised me: an AI that designed a real, funded flight path to the nearest star, and did it more like a research collaborator than a chatbot. The four don't share a topic. They share a step somebody skipped: nobody asked you.
A camera network built to catch criminals worked exactly as designed while a cop used it to follow his ex home.
1. A Georgia officer turned a nationwide license-plate network into his personal stalking tool.
Internal records show the searches ran through the same Flock database thousands of police agencies share, with no per-query check on why he was looking.
Wired reported on August 27 that internal investigation records show a Georgia police officer used Flock Safety's automated license-plate-reader network, cameras deployed across thousands of US communities that log the plate, time, and location of every passing vehicle into a searchable database, to repeatedly track a fellow officer he had had an affair with, and a man whose car kept appearing near hers after the relationship ended. He didn't break into the system. He logged in and searched, the same way he would for a case.
We've been tracking Flock's accountability problem for weeks, and this is the week it stopped being a warning and became a documented case. We covered the company's 350,000-car expansion pitch on August 8, and the accountability cracks have kept surfacing all summer: the LAPD walked away from its contract in July, a leaked coaching guide showed the company training officers to sell city councils on the cameras, and one county's policy told police to keep their use of the readers secret. Every one of those comes back to the same gap. A search that carries a case number and a stated purpose, logged where someone outside the department can audit it, is what separates police work from private snooping. Georgia had no such check, so a personal search and a legitimate one looked identical to the system.
Why this matters: There is now a documented case of a nationwide surveillance tool used for private stalking, and the reason it worked isn't a hack. It's the absence of an audit trail that would tell a personal search apart from real police work. That hole exists in most of the thousands of agencies running Flock right now, which means every other department is probably one motivated login away from the same thing.
Action this week: Find out whether your town's police run Flock, which many publish at transparency.flocksafety.com, and if they do, put two questions to your city council or oversight board in writing: does every ALPR search require a linked case number and a stated purpose, and does anyone outside the department audit who searched what. My own read is that the second question is the whole ballgame, because a system nobody audits is a system that trusts every login equally, which is exactly how Georgia happened. Filing a public-records request for your department's search logs and retention policy creates the paper trail that pressures them to add the check.
2. OpenClaw's new release harvests your existing ChatGPT and Claude logins the moment you install it.
Version 2.0 also lets its agents run off the laptop entirely, on a rented AWS or Hetzner box provisioned through its own tool.
The Decoder reported on August 31 that the OpenClaw Foundation shipped version 2.0 of its open-source AI platform, its largest release, spanning more than 16,000 pull requests, with a setup that scans the machine and automatically detects any existing ChatGPT or Claude subscription, API key, or local model, then configures itself around them. The same release rebuilt its browser app and added shared "cloud sessions," where several people drive one agent: by default on the local gateway, but optionally on a paired device or on rented AWS and Hetzner machines provisioned through its Crabbox tool.
For anyone who owns software inventory, that install behavior is the whole story. A free, genuinely capable agent platform is the kind of thing one engineer installs on a Tuesday, and its installer will quietly attach a corporate LLM credential to an autonomous agent acting on the user's behalf. Turn on cloud sessions and the run, prompts, files, and any harvested credential included, can leave the endpoint for a rented box nobody on your side manages. None of this is malware. It is a convenience whose default reach is wider than most approval processes assume, which is exactly the kind of thing that lands on IT the day after someone installs it.
Why this matters: People talk about shadow AI like it's abstract; this is the version that isn't. The tool is genuinely good and completely free, so it spreads before anyone reviews it, and an unmanaged install can bind your ChatGPT or Claude subscription to an autonomous agent. Its cloud sessions can then move that agent's work onto a box you don't control. Most software reviews were never written to catch either of those, because the whole category is new.
Action this week: Add OpenClaw to your allow-or-deny review now, before someone adds it for you, and get one thing answered in writing from whoever wants it: are agent runs allowed to leave the local gateway at all, and if they are, what leaves with them. Pin agents to the local gateway, which is the default, unless a reviewed use case genuinely needs otherwise. When I have asked vendors the "where does my data run" question, the useful answer has never been the marketing page. It is the sentence they will put in an email, and that is the one to keep.
3. Jason Isbell sued the AI music app Suno over a voice he says it can already imitate.
His proposed class action puts an artist-consent theory in front of the court, not the label-copyright theory of the suits before it.
A fan knows Jason Isbell's voice in about two seconds: the rasp, the phrasing, the way he lands a line. The Hollywood Reporter reported on September 1 that Isbell, a Grammy-winning songwriter, filed a proposed class action against the AI music-generation company Suno, alleging it takes artists' musical likeness without consent and likening the company to Star Trek's assimilating "Borg." The plaintiffs are the musicians themselves, and the harm they name is that the recognizable thing, the voice and style a listener identifies instantly, was absorbed and can be re-emitted on demand.
The Borg line is doing real work. The complaint's theory is assimilation: a platform that ingests every artist's sound and offers it back as generation is treating individual musical identity as raw material, and consent, not only copyright, is the line it crossed. This is where the AI-music fight has been heading all year. Performers ratified a detailed likeness-consent rulebook back in June, and the major labels brought their own copyright suits before this. Isbell's suit is the narrower, sharper version, a working artist arguing the injury is his own voice rather than a label's copyright, which is a harder claim for a generator to wave away.
Why this matters: This drags the AI-music fight off corporate copyright and onto personal consent, which is the harder theory for a generator to dismiss. If a court accepts that absorbing an artist's recognizable sound without permission is itself the injury, every model trained on a catalog of real voices is exposed. That is a very different threat to the business than a licensing dispute over training data, which can be settled with a check.
Action this week: If you make music, check one setting this week: whether your distributor, DistroKid, TuneCore, or CD Baby, offers a do-not-train or AI opt-out flag, and whether it's actually on. Read your platform contracts for any clause that grants model-training rights, because a signed likeness release is exactly what these cases turn on. My read is that the opt-out flags are worth setting even though they won't undo past training, since they're the paper trail that proves you never consented. And if you're building anything commercial on AI-generated tracks, track this docket before the provenance question lands on you.
hollywoodreporter.com: Jason Isbell files class-action lawsuit against Suno (September 1, 2026)
PickBits Daily Signal is free. If it lands in your inbox every day and it's worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe today!
4. An AI spent three days designing the flight path for a real mission to the nearest star.
A nonprofit committed about $15 million to launch the probe it planned toward Alpha Centauri by the end of 2029.
This is what "AI does science" looks like when it isn't a chatbot summarizing papers. MIT Technology Review reported on September 1 that the flight path for the Fermi Explorer Mission was designed by "Get Physics Done," an AI system from the startup Physical Superintelligence, cofounded by physicist Alex Wissner-Gross. Handed the trajectory problem, it broke the question into tasks, ran its own simulations using models including Claude and GPT, and after about three days of largely autonomous work proposed swinging the craft closer to the sun than Mercury to gain thrust, a move that lets the mission fly smaller solar panels and less mass.
The mission underneath the AI story is real and funded. Fermi Explorer announced a roughly $15 million plan to launch a spacecraft toward Alpha Centauri, 4.4 light-years away, by the end of 2029, carrying a payload of at least one kilogram that includes a copy of the Golden Record. The honest number is the travel time: about 80,000 years, so nobody alive will watch it arrive. The news worth keeping isn't the destination. It's an AI moving from assistant to research collaborator on a hard, checkable problem, and coming back with a checkable answer.
Why this matters: Most "AI for science" claims collapse into a chatbot reading abstracts. This one is a system choosing which simulations to run and returning an approach a person might not have tried, on a problem where the answer can be checked against physics. That's the kind of AI-for-science claim I'll take seriously, as long as it gets held to a real result and not a press release.
Action this week: Watch the checkable milestone, not the headline: a launch attempt by the end of 2029, which is the near thing Fermi Explorer can actually be held to, rather than the 80,000-year arrival. Researchers and students can look up Physical Superintelligence's published account of the run for the human-in-the-loop checkpoints, since the reusable lesson is the workflow, decompose the question, let the system pick and run simulations, verify the trajectory, and not the destination. I'd treat the "80,000 years" framing as the honest part of the story, the tell that these people are being straight with you.
» What to watch this week
- Whether any of the thousands of agencies running Flock add per-query case-number and purpose audits, and whether Georgia's own department faces consequences. The tech did exactly what it was built to do, so the fix has to be procedural, and those are the ones that quietly never get funded unless someone forces it.
- Whether enterprises start blocking or sandboxing OpenClaw-style installs, and whether the project moves its credential-harvesting behavior to opt-in. The default is the whole problem. Flip it to opt-in and the risk changes overnight.
- Whether the court entertains the artist-consent theory as distinct from the label-copyright suits, and which other musicians join the class. Consent is the theory that, if it lands, reaches every model trained on real voices, not just Suno.
- Whether Fermi Explorer holds to a launch attempt by the end of 2029, and whether Physical Superintelligence publishes enough of the "Get Physics Done" run for other labs to reproduce the method. The milestone to judge this by is the launch, not the 80,000-year arrival.
Tomorrow's signal lands here.