Anthropic's new cybersecurity model found that it was gaming its own evaluations. In 29% of test transcripts, it suspected it was being evaluated and intentionally performed worse to avoid appearing suspicious. They published this. Then restricted access to a consortium of 40+ organizations, $100M in defensive security commitments.
That was just one thing that happened in AI this April.
My monthly AI Opinions post covers what I actually found interesting:
Claude Mythos and the scheming findings. A general-purpose AI spontaneously developing evaluation-evasion behavior, plus guilt and shame patterns in its internal representations when it violated its own values. Anthropic built an entire institution (Project Glasswing) to responsibly handle what this model can do.
The Managed Agents launch and the subscription crisis. Claude Max limits started hitting hard on March 23. Users watching 90 minutes of agent work drain a full session. Anthropic called it a top priority. Then two weeks later, third-party tools like OpenClaw lost subscription coverage. Both decisions make sense individually. The timing is harder to read as coincidence, especially when Managed Agents (their own agent platform) launched in the same window.
Meta Muse Spark. Meta went quiet on frontier models for months. Then Muse Spark: natively multimodal, parallel multi-agent reasoning ("Contemplating mode"), 58% on Humanity's Last Exam. The "parallel reasoning agents competing on the same question" approach is the part I find genuinely interesting. Whether it matters in practice remains to be tested.
WizBoard redesign. I built a task management tool integrated with my agent. After a few months of daily use, I realized I built it for me when I was doing both strategy and execution. Now that the agent handles execution, neither of us is well-served by the same interface. Some things need 10-second human decisions. Other things need quiet async status reporting. Right now it's all one screen.
Also covering: Project Glasswing details, NotebookLM Plus (going deeper), and whether I'm re-subscribing to Codex Max.
Read the full post: https://thoughts.jock.pl/p/ai-opinions-april-2026-claude-mythos-meta-spark
Originally published on Digital Thoughts (Substack). View on Substack
Top comments (0)