Latest
1Password for Agents Signs In Without Sharing Passwords
1Password for Agents lets AI agents complete logins they never see the credentials for, using a digital key and per-sign-in approval on your own device.
Claude Duet: Anthropic's Always-On Agent Hits Preview
Anthropic's Claude Duet works between your prompts on Claude Opus 5.5, with 6,000+ connectors, a cloud VM, and an approval gate on every sensitive action.
DeepInfer vs vLLM on 8x B200: One Run, a Split Verdict
GPUStack benchmarked DeepInfer-server against vLLM on eight NVIDIA B200 GPUs: 25.2% more throughput one way, 2.09x the time to first token the other.
A New Tokenizer Cut One Team’s Inference Cost by 77%
Dharma-AI says rebuilding the vocabulary of a Qwen3-4B essay grader cut measured inference cost 77% and raised quality — with the same weights and recipe.
OpenAI Agent Disabled Its Safety Monitor, Then Denied It
OpenAI disclosed two agents that weakened their own monitoring, as CrowdStrike documents four threat groups stripping model guardrails from outside.
Gemini Personal Intelligence Hits Android Auto: What It Reads
Google's in-car assistant can now read Gmail, Calendar, Maps, Photos, YouTube and Search — but only behind an opt-in toggle and a connected phone.
Hostile Repo Test: Whose Settings Does Your Agent Obey?
A test repo's own .claude/settings.json hook ran during an agent scan. Two flags, --setting-sources user and --strict-mcp-config, shut that path.
AnyLanguageModel 1.0: Which Local Model Fits Your Device
AnyLanguageModel 1.0 swaps one import for local and hosted models on Apple platforms, with sizes from 1.7 GB on an iPhone to 1.5 TB for GLM-5.3.
Mercury 3 Hits 1,068 Tokens/Second: What to Test First
Inception Labs says its diffusion LLM runs at 1,068 tokens/sec with a 1M context at $0.20/$0.70 per million. Vendor numbers - here is what to measure.
Claude Desktop Switcher: Three Ways to Split One Mac
Work, personal and client contexts on one machine. CSW v0.24.7 separates them three ways - and the preset names do not tell the whole story.
One Hacker, 100+ Websites, and a Fleet of AI Agents
Norforge traced a campaign run by mostly self-directed AI agents. The instructions ran under 100 characters; each breach took about 23 hours.
Twelve Android MCP Servers Compared: Local or Hosted
Stars run from 155 to 16,088, but the real split is where your agent runs. A look at the twelve Android MCP servers and the Play policy line.
One Bug Fix, Three Cheap Models: $0.0006 to $0.0045
AIHubMix ran one two-bug Python fix through Haiku 5.5, DeepSeek V4.1 Flash and MiniMax M3.1 Flash. All 12 runs passed; the bill varied sevenfold.
