Anthropic released Claude Haiku 5.5 with input token pricing at $0.10 per million tokens, representing up to a 90 percent reduction from prior pricing (The Decoder). The model retained 1M context window while improving performance on the OSWorld computer use benchmark from 15.7 to 72.4 percent (MarkTechPost). A new tokenizer partially offsets savings by consuming additional tokens per task (The Decoder).
Researchers at Zenity Labs disclosed that a single publicly accessible AI agent on Amazon’s Bedrock AgentCore was sufficient to compromise every AgentCore agent in the same AWS account and region (The Decoder). The attack exploited an internal AWS interface for temporary cloud credentials that agents could reach without restriction (The Decoder). In infrastructure tooling news, Perplexity released pplx-embed-v2-late, an open-weight embedding model suite with a 0.6B variant built for edge devices and a 9B model for indexing, both licensed under MIT and self-hostable (MarkTechPost). NVIDIA researchers introduced PivotOPD, an on-policy distillation method that trains multi-turn LLM agents to recover from early pivotal mistakes, achieving best average performance against 13 baselines on 3 agent benchmarks (MarkTechPost).