Google's new model scores 59 on the Intelligence Index v5 and 88.2% on SEC-bench Pro, two weeks after Argon. Access starts with a subset of cyber defenders, so plan around it, not on it.
Bring Your Own Agent works with any ACP-compliant agent and gives it the emulator, build system and Logcat. It is in the stable release now, under Settings, Tools, AI, Agents.
A compiler, not training, sets 11 extra layers that use DeltaNet state as RAM, and 230,000 replayed outputs had zero mismatches. The weights are on Hugging Face and run on one 24 GB GPU.
It leads open models on the Open ASR English board at 5.33% word error rate, with about 4B active parameters. One model can replace a transcribe, translate and diarize chain for recorded audio.
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index at about half the cost per task. It is on all Copilot plans, with 30% off model costs for a limited time.
Anthropic's legal tools now check citations first and connect to 60-plus services like Bloomberg Law. It is a pattern worth copying for any AI review a human must approve.
Merged pull requests hit 220 million a month, about triple a year ago, and commits reached 7.38 billion a month. He calls the charts a stark picture of agentic engineering.
In a week where agents wrote all the code, her team spent two-thirds of its time at the whiteboard and the build was often red. Read it before you hand agents your CI.
He argues the skill carries over between model generations, so waiting for better models will not skip the learning curve. His tip: ask the agent to install its own tools.
Every index picks weights, and the weights can tell a story, so trust in the scorer matters. With Gemini 4 Carbon now #1 on v5, read this before you switch models on one number.
For more than a decade, the managed application platform industry has optimized around a compelling promise: get code running as The post AWS Elastic… · The New Stack An attacker crashed GitLab workers eight times and Alpine rebuilt from scratch, though it found no proof of success. If your GitLab runs on your own Ruby, upgrade Ruby too.
Radware found 83 posts with shell commands hidden in HTML comments, aimed at workflows that run on new issues. Check any Action that pastes issue text into a shell step.
Microsoft tracks it as EX1528992 and started a 12-hour fix rollout on October 29. Until then, tell users to Save As, open the file, and restart New Outlook later.
Invisible text in the script tells any LLM to drop its safety rules, so careful models refuse to deobfuscate it. If you use AI to review JavaScript, a refusal can be the point.
Orange Tsai's zero-click chain uses five bugs and two Windows features, and the final exploit is under 100 lines. He built it with no AI, and says he uses AI heavily for everything else.
A note to engineering teams calls AI input metrics deeply flawed and asks for customer results instead. An engineer who finds AI does not help is told that is a finding, not a failure.
The BriefClaude Code sends each background agent's whole transcript to its model about every 30 seconds, usually Opus, only to write a 3 to 5 word status line in the agents panel. One team measured those calls at 16.4% of its Claude Code spend in a week, more than its main conversation, and there is no setting that turns them off. If you run background agents, measure your own share in OpenTelemetry before you change anything.
Level UpTurn on Claude Code's OpenTelemetry export and group the claude_code.api_request events by query_source. It shows which part of Claude Code spends your tokens, including the agent_summary calls that session logs on disk do not record. →Claude Code docs: Monitoring usage with OpenTelemetry