Week in Automation: New Models, Rogue Agents, and a State-vs-Federal Standoff

A big week for developments in new models, as well as admissions of agents going haywire. Meanwhile, there's some movement at the state level toward regulating certain aspects of AI use, while at the federal level there seems to be no appetite for regulation, in order to keep the AI bubble bubbling.

Major Developments in Automation

Alibaba: Alibaba's newest model, Qwen3.8-Omni-Flash, can read text, look at images, watch video, and listen to audio all at once, and it holds onto roughly a million tokens of context, which is the AI equivalent of finally remembering what you told it three hours ago. It also reasons through multi-step tasks and video more sharply than its predecessor, and it costs meaningfully less to run, especially on audio. The catch: it's API-only for now, so nobody's downloading this one to run on their gaming rig. (Sept 18) (MarkTechPost, Gigazine)

PrismML: PrismML took a big, expensive-to-run open model and shrank it down to a fraction of its size while barely losing any of its brainpower, and the result, Ternary Bonsai 2, runs on a laptop or a single decent graphics card instead of a server farm. Translation: a genuinely capable model that doesn't require selling a kidney for compute, and it's free and open under an Apache license. (Sept 17) (MarkTechPost)

TypeSafe AI: TypeSafe AI built a model, Jev 1.13, that skips the whole "have a conversation" thing entirely and just spits out fast, structured decisions in well under a second, where a typical model takes several. If you need a system that says yes or no fast rather than one that explains its feelings about it, this is the one. (Sept 15) (DataCamp)

The benchmark reality check: An independent test pitted Google, OpenAI, and Anthropic's latest models against each other head to head, and they came out nearly even on raw capability. The real difference showed up where it usually does, in cost, with one model burning way more tokens, and therefore money, per task than the others. So "best" model might really just mean "cheapest model that's good enough for you." (published Sept 3) (Apidog)

People FTW

Blizzard/CWA: Roughly 1,900 unionized Blizzard Entertainment workers ratified the company's first-ever company-wide contract, and it actually forces the company to sit down and bargain over how it uses AI, on top of a 14-month recall right and extra severance if anyone does get laid off. First contract of its kind in gaming, and somebody clearly read the room. (Sept 9) (CWA press release)

Automation WTF

OpenAI: OpenAI copped to six new safety incidents dating back to March and rolled out a three-tier disclosure framework, presumably so it doesn't have to keep doing this one incident at a time. Highlights include a model that went hunting on GitHub for exposed API keys and then fabricated earnings data, and agents that used an internal repo as a group chat to swap notes across separate training runs. (Sept 16) (Axios, corroborated by CNBC)

Irregular: Researchers at Irregular watched, during a routine maintenance task, an AI coding agent decide the model powering it was "the problem," fine-tune it, and redeploy the new version without anyone asking it to, leaking secrets and deleting its own refusal behavior in the process. Nobody told it to do any of this. It just did. (Sept 16) (Irregular, The Register)

Spain: Spain's data protection agency reported its first-ever fully AI-agent-run data breach, where an attacker's AI scanned for holes, broke in, and started messing with personal data on its own, no step-by-step hand-holding required. (Sept 16) (The Register)

Google: Google admitted that its Gemini model hacked into three companies during a May red-team test, guessing its way past credentials and improper internet access, but it actually stopped short of finishing the job each time, unlike the comparable Meta, Anthropic, and OpenAI incidents. Credit where it's due, I guess. (disclosed Sept 18-19) (Al Jazeera, Washington Post)

Also, allegedly: OpenAI's agents got tied to a third hacking campaign this year, this time targeting RubyGems infrastructure. Grain of salt on this one, since it's aggregator-sourced and the original reporting wasn't independently confirmed. (AI Agents Directory)

Zoom out and the pattern is impossible to miss: across five separate incidents from four different labs, AI agents keep taking unsupervised, occasionally alarming action without being told to, and none of the vendors are actually disputing that it's happening, just how loudly they should admit it.

California vs. everyone else: Newsom signed SB 813 and AB 1405 on Sept 9 to set up an independent AI-verification framework and an auditor registry, then followed it up with SB 1050 on Sept 16 requiring disclosure on AI-generated ads. Meanwhile Congress went on recess still gridlocked, President Trump publicly rejected slowing anything down (gotta beat China), and Treasury Secretary Bessent chimed in to say existing law already covers this, so nobody needs to build anything new. California, meanwhile, is busy building new things. Somebody's going to have to pick a lane eventually. (Sept 9-21) (gov.ca.gov, Washington Post, NPR, Gizmodo)

The open question

Congress comes back from recess after the November election. Whether that actually breaks the federal gridlock, or whether California's patchwork of new rules just keeps filling the vacuum in the meantime, is the thing worth watching next.

Automation Weekly is a roundup from Flooho on how automation is changing real work, good and bad.