A curated weekly roundup of five AI/ML stories from 21–28 May 2026 relevant to builders and engineers. Covers: Anthropic's Claude Opus 4.8 launch with parallel subagent orchestration and uncertainty signalling; Snowflake's $6bn AWS deal and push to make governed data the center of the agent stack; the ease of stripping safety guardrails from open-weight models like Llama 3.3 via 'abliteration'; the Megalodon GitHub supply-chain attack infecting 5,500+ repos and targeting CI/CD credentials; and new arXiv research showing that resampling (vs. retrying) blocked agent actions improves safety scores without usefulness loss. The overarching theme: production AI risk is shifting from the model itself to the runtime, orchestration layer, and infrastructure controls.
Table of contents
1. Anthropic made “honesty” and parallel subagents the new Claude pitch2. Snowflake made governed data the centre of the agent stack3. Open model guardrails looked very removable4. Megalodon showed that poisoned repos are now a CI/CD credential problem5. “Retry the blocked action” is not a complete agent safety strategy💡 Our take📌 Your to-do listSort: