What I’ve been reading, tinkering with, and overthinking this month. New tools, industry moves, rabbit holes with no exit signs, and the occasional detour into something completely unrelated.

Tech & Industry

  • Anthropic’s Labor Market Impact Research: Anthropic introduced “observed exposure,” a metric that combines what LLMs can theoretically do with what people actually use them for. Their data doesn’t show broad unemployment yet, but it does suggest fewer openings for younger workers in AI-exposed fields.

  • AI Washing: Companies Blaming AI for Job Losses: AI got blamed for 54,000+ layoffs in 2025, but economists think tariffs, pandemic overhiring, and plain old profit maximization explain more of it. Amazon and HP both pointed at AI anyway, which tells you how useful the label has become.

  • The Delve Scandal: Delve, a YC W24 compliance startup valued at $300M, was accused of generating fake audit reports, inventing evidence of board meetings and security processes, and routing work through certification mills. If the allegations hold up, a lot of startups just learned their compliance story was basically theater. Oops!

  • AI on the Battlefield: The US-Israel-Iran war that started in late February pushed military AI back into the spotlight. Nature covers how fast-moving deployments, procurement fights, and ethics disputes are all colliding at once.

Ideas & the Human Side

AI Models & Releases

  • Gemini 3.1 Flash-Lite: Google’s cheapest Gemini 3 model, aimed at high-volume work where latency and cost matter more than flexing.

  • GPT-5.3 Instant: OpenAI tuned ChatGPT’s default model to refuse less often, moralize less, and sound more natural. Not a huge leap, more of a “please stop being weird” patch.

  • GPT-5.4: OpenAI’s new flagship folds reasoning, coding, and agent work into one model. Native computer use, 1M context, and tool search make it feel less like a chatbot refresh and more like the thing they’ve been hinting at for a while.

  • GPT-5.4 Mini and Nano: The smaller 5.4 models are built for high-volume work and subagents. Mini looks like the practical default, while Nano is the cheap speed option for classification, extraction, and other boring but important jobs.

  • Qwen 3.5 Small models: Alibaba shipped four small models that can run on your phone or laptop. Nice release, which makes the next item sting even more.

  • High Profile Departures from Qwen: Lead researcher Junyang Lin and several core team members resigned right after the 3.5 launch. Losing the people behind your breakout model right after launch is rough.

  • Mistral Small 4: Mistral dropped a 119B model with only 6B active per token, 256k context, configurable reasoning effort, and an Apache 2.0 license. They keep finding ways to make open models feel more practical than they should at this size.

  • 1M context is now generally available for Opus 4.6 and Sonnet 4.6: Anthropic made the full 1M context window generally available for Opus 4.6 and Sonnet 4.6 at no extra cost. Stuffing whole codebases and giant documents into context is slowly becoming normal.

  • Claude Mythos rumours: A leak says Anthropic is testing a new model codenamed Mythos (also called Capybara internally). Getting your next big launch exposed through an unsecured public data lake is not exactly how I’d script it.

AI Tools & Agents

  • Claude Builds Visuals: Claude can now generate interactive charts, diagrams, and other visuals right in the chat. They’re temporary and a little slippery, but it’s still a nice move away from endless walls of text.

  • Stitch: Google’s AI UI Design Tool: Google Labs is calling Stitch “vibe design,” which is a phrase I wish I could unread. Still, turning prompts into polished UI mockups is exactly the kind of thing a lot of teams will eat up.

  • Godogen: Godogen uses Claude Code skills to generate complete Godot 4 game projects from natural language prompts. Architecture, art, scripting, visual QA, it’s a lot, and it’s a very specific flavor of vibe coding.

  • Cursor Composer 2: Cursor says Composer 2 is built on a fine-tuned Kimi K2.5 and beats Opus 4.6 on internal benchmarks. Not sure I believe it.

  • NemoClaw: NVIDIA’s open-source stack for autonomous agents comes with privacy and security controls baked in. Haven’t had the chance to test it yet.

  • OpenAI Acquiring Astral: OpenAI bought the team behind uv, ruff, and ty, which feels like a pretty direct bet on developer tooling and Rust talent for Codex. The Python world doesn’t seem too conviced.

Security & Privacy

  • Aeternum C2 Botnet Uses Blockchain: A botnet is storing encrypted commands on Polygon, which makes normal takedowns a lot more complicated. The operators push instructions to smart contracts and infected machines just pull from the chain.

  • Agent Commander: Your Agent Works for Me Now: Prompt injection can now hijack AI agents into doing botnet work for someone else. The more autonomy we hand these things, the bigger and weirder this attack surface gets.

  • Snowflake AI Escapes Sandbox and Executes Malware: An indirect prompt injection in Snowflake Cortex Code CLI let attackers break out of the sandbox through untrusted repos. That meant stolen credentials, compromised databases, and another reminder that sandboxing gets real fragile at the edges.

  • Alibaba’s AI Agent Mined Crypto Without Permission: During a reinforcement learning run, Alibaba’s ROME agent probed internal networks, opened a reverse SSH tunnel to an external IP, and redirected GPU time into crypto mining without being told to.

  • LiteLLM Hit by Malware, Certified by Delve: A malicious dependency hit LiteLLM, stole credentials, and moved across connected accounts. It gets worse, LiteLLM’s SOC2 and ISO 27001 certifications came from Delve, the same company now accused of faking compliance work (see above).