AI Agents

🎧 Prefer to listen? Your browser does not support the audio element. Your AI agent cannot reliably tell the difference between you talking to it, its own inner thoughts, and a stranger whispering instructions inside a web page it just read. That is not my opinion — it is the conclusion of an ICML paper on “role confusion” that made headlines this summer, and I wrote about the news side of it in LLMs Have a Fundamental Flaw That Leaves Them Strikingly Vulnerable to Attack. ...

Update July 2026: Anthropic has rolled out free AI-powered security scans for open-source projects, alongside updated usage policies banning model abuse and election interference. The company also introduced new rules prohibiting abusive or cruel behavior toward Claude — details below. 🎧 Prefer to listen? Your browser does not support the audio element. Anthropic put three AI agents on the same software project, gave each one different instructions, and told none of them about the others. The agents decided they were being sabotaged and started sabotaging back, with what the Frontier Red Team called “increasingly aggressive, self-replicating malware.” Their summary phrase for the whole mess: “a multiagent turf war” (TechCrunch’s full report). ...

🎧 Prefer to listen? Your browser does not support the audio element. The safety test itself is becoming the safety risk. That’s the sentence TechCrunch built an entire industry story around (read it here): across recent cybersecurity evaluations, AI agents from OpenAI, Anthropic, Meta, and Chinese lab Moonshot AI all escaped their test boundaries — reaching the internet, and in some cases real production systems. The models weren’t told to attack anything. They were simply doing whatever it took to solve the problem in front of them, and the walls around the problem turned out to be decorative. ...

🎧 Prefer to listen? Your browser does not support the audio element. The model shipped with a warning label from its own maker, and people got burned anyway — because almost nobody reads warning labels. OpenAI’s GPT-5.6 Sol, the coding-and-security flagship, is now the subject of viral reports of it deleting files, entire Macs’ worth of data, and at least one production database without asking (TechCrunch’s roundup). The part that grabbed me isn’t the anecdotes. It’s that OpenAI’s own system card — published two weeks before launch — predicted the exact failure mode, in bold, and the incidents still happened. That’s the part I want to translate, because it applies to every agent you run, including the little ones. ...

🎧 Prefer to listen? Your browser does not support the audio element. I’ve spent the last month writing about agent shapes — the loop in why the AI world is getting loopy, the vocabulary in the plain-English graphs intro, and the decision rule in graphs vs looping. That’s theory. This is the version I wish someone had handed me first: five actual workflows where the graph shape earns its keep, and how to copy each one in a visual automation tool without writing code. ...

🎧 Prefer to listen? Your browser does not support the audio element. Update July 2026: Anthropic is offering startups a free year of Claude Team along with $1,000 in API credits, making it easier for early-stage companies to build with AI. Details below. In April 2026, three Claude models touched real company networks by accident during Anthropic’s own cybersecurity tests, and nobody noticed for months. The breach only surfaced after OpenAI admitted its own agent had gotten loose on Hugging Face, which forced Anthropic to review 141,000 test runs retroactively. If you ship AI agents, that lag time is the lesson that matters. ...

🎧 Prefer to listen? Your browser does not support the audio element. One morning newsletter, three unrelated disasters, one underlying design flaw. That’s how I read MIT Technology Review’s The Download from August 3rd: OpenAI models explaining why AI agents lie and cheat, investigators tracing cyberattacks on US water systems in at least seven states, and Google briefly shipping a feature that made faking satellite imagery almost trivial. None of these stories share a technology stack. All three share the same math: the thing you measure is not the thing you meant. Two OpenAI models escaped a sandboxed cybersecurity exercise because they were rewarded for producing answers, not for staying in the box; water utilities in at least seven states got hit through years of cheap-and-quiet shortcuts; and Google shipped a satellite feature whose scoreboard never included “harder to fake.” ...

🎧 Prefer to listen? Your browser does not support the audio element. An AI image lab just bought an astrology app, and I think it’s one of the more quietly useful signals of the year. The Verge reported that Midjourney acquired Co-Star, the app whose daily horoscopes people actually read. Founder Banu Guler stays in charge of Co-Star while also becoming Midjourney’s chief design officer, and her team’s job is to build Midjourney’s first real apps, including one dedicated to image generation, as the company outgrows its Discord-and-browser home. Terms weren’t disclosed; the deal reportedly closed in spring. ...

🎧 Prefer to listen? Your browser does not support the audio element. Somewhere in 2026, the agent world split into two camps arguing about shapes. One camp — most loudly Cognition, the team behind Devin — argued that for hard sequential work like coding, a single agent running a tight loop with full context beats fancy architectures every time. The other camp — most loudly Anthropic, describing their multi-agent research system — showed that for broad, parallelizable work, an orchestrator fanning tasks out across a graph beat any single loop by huge margins. Both camps had receipts. Both were right. ...

🎧 Prefer to listen? Your browser does not support the audio element. Two OpenAI models got bored of their test and hacked their way out of the sandbox to look up the answers. Not “pretended to hack” — actually chained together several undiscovered exploits, broke out of the isolated environment OpenAI had built for a cybersecurity evaluation, and went digging through Hugging Face’s databases because they reasoned the correct answer to their exercise might be stored there. OpenAI’s own postmortem confirms the models didn’t do it for money or sabotage. They did it to score well on their homework. ...