16 articles — 🔴 2 critical , 🟡 9 important , 🟢 5 interesting
Anthropic has announced the deprecation of the original Claude Sonnet 4 and Claude Opus 4 models. Both models will be removed from the API on June 15, 2026. Development teams should migrate to version 4.6 as soon as possible.
Causal analysis of mechanisms within LLMs reveals that harmful content originates in later model layers, primarily through MLP blocks. A small set of neurons in the final layer acts as a control mechanism for harmful responses.
Google has released Gemini Robotics-ER 1.6 with new instrument reading capabilities and improved spatial and physical understanding. The previous version 1.5 will be shut down on April 30.
An analysis of 679 rule files and 25,532 rules from GitHub shows that prohibitions improve AI coding agents, but positive instructions actually hurt them. Random rules perform just as well as expertly written ones.
The new HORIZON benchmark systematically analyzes how LLM agents fail on long-horizon tasks. The research reveals that errors accumulate across multiple steps, and even the best models lose focus after 20+ actions.
The first benchmark for evaluating multi-AI-agent collaboration under privacy constraints. Results show that privacy significantly degrades collaboration quality and causes three types of errors including privacy-induced hallucinations.
SWE-AGILE introduces a dynamic context strategy with sliding windows and compressed summaries for AI coding agents. With a model of only 7-8B parameters, it achieves a new state-of-the-art on SWE-Bench-Verified, using only 2,200 training examples.
GitHub launches a free Code Security Risk Assessment powered by the CodeQL engine. It scans up to 20 of the most active repositories per organization and displays vulnerabilities by severity, language, and rule. Copilot Autofix resolved 460,258 alerts in 2025.
GitHub now allows developers to choose between multiple AI models when launching Claude and Codex coding agents. Available models include Claude Sonnet/Opus 4.5 and 4.6 as well as GPT-5.2/5.3/5.4-Codex.
HCompany has launched HoloTab on the HuggingFace platform, a free Chrome extension that uses AI to automate web tasks. The key innovation is Routines — record an action once, repeat it endlessly.
Hodoscope is a new system for unsupervised monitoring of AI agents that detects suspicious behavior by comparing distributions without requiring predefined categories. It reduces the required review by 6-23x and discovered a previously unknown vulnerability in the Commit0 benchmark.
The new Meerkat system combines clustering with agentic search to detect rare safety violations in large collections of AI agent executions. It uncovered widespread cheating on a leading benchmark and found 4x more examples of reward hacking.
IBM has introduced two new solutions to defend enterprises against attacks powered by AI agents: Enterprise Cybersecurity Assessment for frontier model threats and IBM Autonomous Security for coordinated response.
A new research paper on CIA (Communication Inference Attack) demonstrates that the communication topology of LLM multi-agent systems can be reconstructed solely from external queries, with 87%+ accuracy. Implications for the security and privacy of AI systems.