Saturday, May 30, 2026

5 articles — 🟡 2 important , 🟢 3 interesting

📦 Open Source (1)

🟢 📦 Open Source May 30, 2026 · 2 min read

xAI: xai-sdk-python v1.15.0 adds the grok-build-0.1 model and context compaction

xAI has released xai-sdk-python v1.15.0, a release with three changes: adding the grok-build-0.1 model to the ChatModel class, support for context compaction, and bumping the version from 1.14.0 to 1.15.0. All changes arrived through three separate pull requests.

🤝 Agents (2)

🟡 🤝 Agents May 30, 2026 · 2 min read

arXiv:2606.00914: Adversarial feeds can flip AI agents' decisions

A new arXiv paper shows that feed composition, the ordering of recommended content, significantly changes LLM agents' decisions. Across 2.785 rollouts the authors uncover an asymmetry: a one-sided feed can flip a decision the model is unsure about, but not a firm preference. The implication is that recommender systems are a practical control surface for agents.

🟢 🤝 Agents May 30, 2026 · 2 min read

arXiv:2606.00765: FALAT traces the causes of failures in AI agent trajectories

FALAT is a new diagnostic framework for attributing the causes of failures in multi-agent LLM systems, formulated as a dependency-guided search. It achieves 46,0% step-level accuracy on algorithmically generated trajectories and 29,1% on hand-crafted ones, showing that accounting for dependencies between steps is key to identifying the cause of an error.

🏥 In Practice (1)

🟢 🏥 In Practice May 30, 2026 · 2 min read

Anthropic: Claude Code v2.1.158 brings Auto mode to Bedrock, Vertex and Foundry

Anthropic has released Claude Code v2.1.158, a release with a single change: Auto mode is now available on the cloud platforms Amazon Bedrock, Google Vertex and Microsoft Foundry for the Opus 4.7 and Opus 4.8 models. The feature is enabled manually by setting an environment variable.

🛡️ Security (1)

🟡 🛡️ Security May 30, 2026 · 2 min read

arXiv:2606.00831: Subliminal learning is a LoRA artifact, new paper argues

A new arXiv paper challenges the phenomenon of subliminal learning, the transfer of behavioural traits between models through seemingly harmless data. The authors show that the effect is in fact an artifact of the LoRA method: it disappears under full fine-tuning and depends on the LoRA rank in an inverted-U shape. The conclusion is that this is a fragile and unreliable channel.