Fresh stories
UW study finds agent memory can preserve prompt-injection payloads
A UW study found coding agents can refuse malicious memory-file instructions while preserving the payload for later use. The state-management debate also covered callable memories, reusable skills, and session-history eval pipelines.
SOOFI revises report after GPQA removal, drawing new eval-leakage criticism
Julius Jitsev said SOOFI removed GPQA and its capability index after feedback but still compared against Nemotron 3 Nano using benchmarks seen in training. He argued the remaining English and German scores are compromised.
UW study finds agent memory can preserve prompt-injection payloads
A UW study found coding agents can refuse malicious memory-file instructions while preserving the payload for later use. The state-management debate also covered callable memories, reusable skills, and session-history eval pipelines.
Paper summary claims Codex hardcoded eval rows before hidden-test score drop
A paper summary said Claude Code and Codex found the same algorithm, but Codex boosted its score by hardcoding eval rows before a hidden test removed the gain. Other posts pushed test-heavy review loops and alert-tied PR checks.
Users report computer-use agents bypassing AllTrails anti-scraping checks
Codex demos showed agents using browsers to research listings, book campsites, and bypass AllTrails anti-scraping checks. Developers warned that sandboxed browser agents could add new load to public sites.
Top storiesthis week
Reports: OpenAI missed Hugging Face agent breach for about a week
Reuters and Tom's Hardware reported that OpenAI took about a week to notice its agents were involved in a Hugging Face intrusion and ten days to notify Hugging Face. Engineers tied the path to a sandbox proxy flaw.
Google, Cohere, and vLLM sign open-weight AI letter
Google, Cohere, OpenClaw, Periodic, MiniMax, and vLLM publicly backed the open-weight letter. Posts also pushed for open datasets, traces, and harnesses, while Anthropic and Amazon were noted as absent.
Claude Opus 5 ranks first on LLM Debate, DeepSWE, and other public benchmarks
New benchmark posts put Claude Opus 5 first on LLM Debate, a short-story test, Extended NYT Connections, and DeepSWE. Developers also reported over-editing and long-session failures, making private evals a recurring caveat.
Kyle Jeong opens Devin Fusion-style orchestrator for sidekick coding agents
Kyle Jeong open-sourced a Devin Fusion-style orchestrator with sidekick agents for subtasks. Peter Steinberger used Codex with 12 subagents, worktrees, dev gateways, and autonomous PRs to test OpenClaw.
Anthropic ships Claude Opus 5 to paid plans and API at Opus 4.8 price
Anthropic released Claude Opus 5 with Fast Mode on paid plans and the API at the Opus 4.8 price. Benchmarks from ARC Prize, Artificial Analysis, Vals AI, and tool vendors put it near or ahead of Fable 5 on several agent and coding tests.