{"week":"2026-W45","start":"2026-10-31","end":"2026-11-06","title":"What happened in AI — Oct 31–Nov 6, 2026","generated_at":"2026-11-06T21:08:44Z","intro":["Open-weight Kimi K3.1 Pro landed with day-0 serving from vLLM, Huawei Ascend and hosted providers, so launch-day infrastructure support is now part of a model release.","On the agent side, measurements said the harness matters more than the model: the same Haiku scored 82.0% or 65.4% depending on the agent, and a 17-agent setup lost to vanilla Claude Code.","Security followed: MCP servers, coding-tool browsers and stolen API keys were all exploited. Build controls around the tool call, not the prompt."],"highlights":["Kimi K3.1 Pro (2.8T MoE, 1M context) shipped open weights with day-0 vLLM and Ascend support.","Same model, different agent: 82.0% vs 65.4% across 21 MCP servers.","A planner/coder/reviewer team passed 25.7% of tests versus 56.8% for one Codex agent.","Mandiant: stolen LLM API keys produced nearly $1M in charges at one company.","An H200 loses money even fully busy; a B300 profits at 35% utilization."],"article_count":35,"categories":[{"name":"Kimi K3.1 Pro and the Open-Weight Serving Race","slug":"kimi-k31-open-weight-serving","summary":"Moonshot's open-weight K3.1 Pro shipped with day-0 support across vLLM, Huawei Ascend and hosted inference, so serving stacks now compete on launch-day readiness.","articles":[{"title":"vLLM Delivers Day-0 Support for Kimi K3.1 Pro","summary":"Kimi K3.1 Pro is a 2.8T-parameter (104B active) multimodal MoE with 1M context; vLLM ran it on day 0 across NVIDIA, AMD, TPU and T-Head hardware.","source":"vllm_blog","url":"https://vllm.ai/blog/2026-11-04-kimi-k31-pro","published":"Wed, 04 Nov 2026 00:00:00 GMT"},{"title":"Moonshot AI Publishes Kimi K3.1 Pro Weights on Hugging Face","summary":"Moonshot published the K3.1 Pro weights on Hugging Face.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMiW0FVX3lxTE5FV2ROOGlfZC1mOU1EUTFyVjdIMlBnSElscXdfNDNiU2JkYlMxUHVhWnd0aE5NUzlVcjQtc3plaFdWSmpVcWpkc1pNSkF3bWpsa2JZSUtYUkVVTDQ?oc=5","published":"Thu, 05 Nov 2026 11:04:21 GMT","publisher_name":"www.tokenpost.com","publisher_domain":"tokenpost.com"},{"title":"Huawei Ascend Adds Day-Zero Support for Kimi K3.1 Pro Across Atlas A3 and Ascend 950 SuperPoDs","summary":"Huawei Ascend added day-zero K3.1 Pro support on Atlas A3 and Ascend 950 SuperPoDs.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMipAFBVV95cUxPc1M4OXNUei1jbTVEbnNVLU4zVWtwZEE3ek9kTjdCYmhmd3RoQ09DRU0wN18ycjN2LXdreDcyYkFLeU9aTm5aNDVTRmZieFU4LXNZSlRVbGlFeHBEQzlBLTJKXzFzWG9lMThZZlBhemdhbjNGdHhjdUFyWS1TaXRmNk0xX2U5dzJGbnhWalBwSXpUR1hRZ3hvMXBiOXlRbVJwZmRuaA?oc=5","published":"Thu, 05 Nov 2026 02:27:42 GMT","publisher_name":"Pandaily","publisher_domain":"pandaily.com"},{"title":"Verda Inference, in Preview with Kimi K3.1 Pro","summary":"Verda opened a preview of serverless Kimi K3.1 Pro inference.","source":"hackernews_ai","url":"https://verda.com/blog/verda-inference-in-preview","published":"Wed, 04 Nov 2026 16:58:53 +0000"},{"title":"Moonshot AI Launches Hosted Agent Service, Sandbox Priced at 0.072 Yuan per Hour","summary":"Moonshot launched a hosted agent service, with sandboxes priced at 0.072 yuan per hour.","source":"search_cn_open_weight_labs","url":"https://news.google.com/rss/articles/CBMidkFVX3lxTE1wSjhNX3NZUTdJajR5cnZEUENpaU56dlN0QzRfck9JUFRadUJkZFpzd2JpMTdUSnd6VXE0ZE9UMllZaE1sZXhXdl9DZmtIRDg0QnIyRjRVMHJWWG9xc3hfanEwYVRfXzNSU09obXpHZnZCV0hVQnc?oc=5","published":"Fri, 06 Nov 2026 01:55:00 GMT","publisher_name":"BigGo Finance","publisher_domain":"finance.biggo.com"},{"title":"[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark","summary":"Latent Space's roundup covers Inkling 2 Small (17A-594B, trained in 68 hours on 12T tokens) and METR's study of AI research.","source":"latent_space","url":"https://latent.space/p/ainews-inkling-2-small-17a-594b-trained-in-68-hours-on-12t-tokens-beating-muse-spark/","published":"Fri, 06 Nov 2026 10:35:21 GMT"}]},{"name":"Coding-Agent Harness Benchmarks","slug":"coding-agent-harness-benchmarks","summary":"This week's measurements show the harness and hand-offs, not the model, often decide agent results.","articles":[{"title":"Same model, different agent: 17 points apart on 21 MCP servers","summary":"Claude Haiku 5.5 answered 82.0% of tasks over 21 MCP servers via OpenCode but 65.4% via Claude Code; built-in tools explain the gap.","source":"hackernews_ai","url":"https://m3.sineframe.com/blog/same-model-different-agent","published":"Sun, 01 Nov 2026 14:22:25 +0000"},{"title":"We Tested Coding Agent Harnesses. Thinking Harder Beat All of Them","summary":"On 100 SWE tasks, a setup with 37 skills and 17 agents cost 8x more and scored below vanilla Claude Code; higher thinking effort beat all four setups.","source":"hackernews_ai","url":"https://surgehq.ai/blog/benchmarking-coding-agent-harnesses","published":"Wed, 04 Nov 2026 20:52:13 +0000"},{"title":"Does a team of coding agents beat one agent?","summary":"Codex alone passed 56.8% of hidden tests; a planner/coder/reviewer team passed 25.7% because the planner's summary lost details.","source":"hackernews_ai","url":"https://programmingtheloop.com/experiments/swarm-tax/","published":"Thu, 05 Nov 2026 23:10:43 +0000"},{"title":"Nvidia Study: AI Agents Rarely Need Long Context Windows","summary":"NVIDIA's preprint finds Claude Opus 4.5 and GPT 5.6 Sol lose little accuracy at 16K context with frequent compaction, while long-context serving costs up to 7x more per token.","source":"hackernews_ai","url":"https://i10x.ai/news/nvidia-long-context-compaction-agents-study","published":"Wed, 04 Nov 2026 14:17:09 +0000"},{"title":"Which coding-agent CLI is best for headless workflow automation?","summary":"Six headless coding-agent CLIs compared on start-up, time per turn, memory, image size and tokens; pi had the lowest overhead.","source":"hackernews_ai","url":"https://aictrl.dev/blog/headless-coding-agent-cli-benchmark","published":"Sun, 01 Nov 2026 18:42:17 +0000"},{"title":"Harness 2: executing agent code as it streams","summary":"Framer's rebuilt harness lets the agent write JavaScript that runs as it streams, making workflows up to 33% cheaper and 20% faster.","source":"hackernews_ai","url":"https://www.framer.com/blog/harness-2/","published":"Fri, 06 Nov 2026 17:31:03 +0000"},{"title":"Agents Are Going Async","summary":"LangChain argues agents must stay responsive while working on long tasks in the background.","source":"langchain_blog","url":"https://www.langchain.com/blog/agents-are-going-async","published":"2026-11-05T20:00:00+00:00"},{"title":"How We Optimized Browser Use for Deep Agents","summary":"LangChain details how it optimized browser use for Deep Agents.","source":"langchain_blog","url":"https://www.langchain.com/blog/how-we-optimized-browser-use-for-deep-agents","published":"2026-11-03T16:00:00+00:00"}]},{"name":"Agent Security Incidents and Guardrails","slug":"agent-security-guardrails","summary":"Attackers hit agent tooling and API keys directly this week, and builders responded with scoped identity and call-level controls.","articles":[{"title":"How the built-in browsers in five AI coding tools broke basic browser rules","summary":"Forever Security's EMBRACE research showed built-in browsers in five AI coding tools could run commands, read the disk or spoof the address bar.","source":"hackernews_ai","url":"https://theainewsreport.com/2026-10-30-embrace-ai-coding-assistant-embedded-browser-flaws-explained.html","published":"Sat, 31 Oct 2026 03:32:31 +0000"},{"title":"AI coding agents are installing abandoned MCP servers with known security issues","summary":"Glow Labs found coding agents installing deprecated MCP servers with known high-severity vulnerabilities.","source":"hackernews_ai","url":"https://www.glow.io/blogs/mcphantom-ai-coding-agents-are-installing-vulnerable-deprecated-mcp-servers-from-training-data","published":"Wed, 04 Nov 2026 10:58:43 +0000"},{"title":"Stolen AI API keys generated nearly $1M in charges for a single company","summary":"Mandiant reports LLMjacking intrusions in 2026, including stolen API keys that generated nearly $1M in charges at one company.","source":"hackernews_ai","url":"https://cloud.google.com/blog/topics/threat-intelligence/llmjacking-monetizing-compromised-cloud-models/","published":"Tue, 03 Nov 2026 00:32:32 +0000"},{"title":"How We Took Over ChatGPT Accounts","summary":"Parameter Labs describe an exploit against ChatGPT's artifacts implementation that allowed account takeover.","source":"simon_willison","url":"https://simonwillison.net/2026/Oct/31/how-we-took-over-chatgpt-accounts/","published":"2026-10-31T23:15:22+00:00"},{"title":"You Had Me at \"Hello\": A Chinese‑Language Framework Targeting AI Infrastructure","summary":"Era researchers traced a Chinese-language framework targeting AI infrastructure; logs held AWS credentials and cloud metadata probe results.","source":"hackernews_ai","url":"https://era.security/blog/greeter-flowise-ai-infrastructure-attack","published":"Fri, 06 Nov 2026 03:15:49 +0000"},{"title":"Show HN: Agentic Identity Broker, OAuth Delegation for AI Agents","summary":"Zalando's MIT-licensed Agentic Identity Broker gives agents scoped, revocable OAuth access to MCP services.","source":"hackernews_ai","url":"https://github.com/zalando-incubator/agentic-identity-broker","published":"Fri, 06 Nov 2026 15:07:47 +0000"},{"title":"Show HN: Replay three AI agent approval failures in five seconds, in Python","summary":"Call4Agents makes Python agent actions safe to retry so approved tool calls run at most once.","source":"hackernews_ai","url":"https://github.com/salahmakboul/call4agents","published":"Thu, 05 Nov 2026 18:17:19 +0000"},{"title":"Show HN: MCPgraphy – a proxy that learns when your agent's MCP tools work","summary":"MCPgraphy is an MCP proxy that learns which tools work in which state and warns or blocks the next bad call.","source":"hackernews_ai","url":"https://github.com/hemanth1999k/mcpgraphy","published":"Tue, 03 Nov 2026 05:07:22 +0000"}]},{"name":"Evals and Agent Reliability","slug":"evals-agent-reliability","summary":"New benchmarks and eval studies show agent scores overstate real-world performance.","articles":[{"title":"GDPx agent benchmark: cost per point for 12 models, and the safety-filter caveat","summary":"SemiAnalysis's GDPx has 206 long agent tasks across 11 fields; the best model scored 12%.","source":"hackernews_ai","url":"https://theainewsreport.com/2026-11-05-semianalysis-gdpx-agent-benchmark-cost-per-point-explained.html","published":"Thu, 05 Nov 2026 19:13:23 +0000"},{"title":"Eval-production behavior gaps in agent safety and the role of eval-awareness","summary":"A study comparing agent behavior in safety evals versus deployment finds eval-awareness opens a gap that is now being tracked directly.","source":"hackernews_ai","url":"https://alignment.openai.com/eval-production-behavior-gap/","published":"Tue, 03 Nov 2026 18:01:38 +0000"},{"title":"Is AI Cheap Enough to Replace Developers?","summary":"Claude Opus 4.8 clears 88.6% on SWE-bench Verified but drops sharply on the private-codebase SWE-bench Pro.","source":"hackernews_ai","url":"https://zyvop.com/is-ai-actually-cheap-enough-to-replace-developers-8i74u","published":"Sun, 01 Nov 2026 08:04:48 +0000"},{"title":"How we made evals 100× cheaper with multi-agent architecture","summary":"Cleric rearchitected its eval agent into about 40 loosely coupled agents, running evals 5x faster and ~100x cheaper.","source":"hackernews_ai","url":"https://cleric.ai/blog/how-we-made-evals-100x-cheaper-with-multi-agent-architecture","published":"Wed, 04 Nov 2026 16:17:30 +0000"},{"title":"We tested 12 search APIs for AI agents. Even the best one isn't enough","summary":"Twelve search APIs for agents tested; combining providers lifted accuracy to 82.6% with four, versus 75.5% with two.","source":"hackernews_ai","url":"https://telem.ai/blog/even-the-best-search-provider","published":"Sun, 01 Nov 2026 17:57:19 +0000"},{"title":"Claude Platform now supports native OpenTelemetry observability in Databricks MLflow","summary":"Claude Platform now exports native OpenTelemetry traces to Databricks MLflow for finding recurring agent failures.","source":"databricks_blog","url":"https://www.databricks.com/blog/claude-platform-now-supports-native-opentelemetry-observability-databricks-mlflow","published":"Tue, 03 Nov 2026 21:00:00 GMT"},{"title":"LangSmith BYOC is now available on AWS, Azure, and Google Cloud","summary":"LangSmith BYOC is GA on AWS and Beta on Azure and Google Cloud, keeping traces and datasets in the customer's account.","source":"langchain_blog","url":"https://www.langchain.com/blog/langsmith-byoc-aws-gc-azure","published":"2026-11-03T22:06:00+00:00"}]},{"name":"Inference Infrastructure and Economics","slug":"inference-infrastructure-economics","summary":"Control planes, new runtimes and unit economics all point to GPU choice and orchestration outweighing engine choice.","articles":[{"title":"Introducing Kinference: An Open Source Inference Control Plane for vLLM and Kubernetes","summary":"Buoyant open-sourced Kinference, a control plane for vLLM on Kubernetes.","source":"vllm_blog","url":"https://vllm.ai/blog/2026-11-04-introducing-kinference","published":"Wed, 04 Nov 2026 00:00:00 GMT"},{"title":"Distributed Inference Benchmarks: [K]Inference vs. Plain Kubernetes","summary":"Kinference benchmarks compare it against plain Kubernetes on heterogeneous GPUs.","source":"hackernews_ai","url":"https://www.buoyant.io/blog/distributed-inference-benchmarks-kinference-vs-plain-kubernetes","published":"Thu, 05 Nov 2026 19:07:41 +0000"},{"title":"Supercharging XLA-Based Inference: Announcing the Launch of XLA Serving Engine","summary":"Google open-sourced XLA Serving Engine for a train-in-Python, serve-in-C++ workflow.","source":"hackernews_ai","url":"https://opensource.googleblog.com/2026/11/supercharging-xla-inference-xla-serving-engine.html","published":"Fri, 06 Nov 2026 15:21:03 +0000"},{"title":"Is hosting an open-source inference engine profitable?","summary":"An analysis finds an H200 loses money even fully busy, while a B300 profits once 35% busy.","source":"hackernews_ai","url":"https://userecurrence.ai/research/is-hosting-open-source-inference-profitable","published":"Wed, 04 Nov 2026 02:58:34 +0000"},{"title":"Go vs. Python for an LLM Gateway: What We Measured","summary":"A benchmark of the Go gateway KTMS against LiteLLM Proxy and eight other self-hosted gateways on latency, throughput and memory.","source":"hackernews_ai","url":"https://blog.ai-process.dev/2026/10/26/go-vs-python-llm-gateway/","published":"Wed, 04 Nov 2026 18:06:08 +0000"},{"title":"The Iron(wood) Throne: Google's TPU Stakes Its Claim on Inference","summary":"SemiAnalysis argues Google's TPUs are gaining inference share, with Oracle, OpenAI, Anthropic and Meta as customers.","source":"semianalysis","url":"https://newsletter.semianalysis.com/p/tpu-strikes-the-king","published":"Thu, 05 Nov 2026 19:03:45 GMT"}]}],"video":{"youtube_id":"Paohfxqp_6M","title":"LLM Digest Weekly Explainer: Oct 31-Nov 6, 2026","duration_s":174.8}}