Run SDK is Vercel's open-source QuickJS sandbox that lets agent-generated JS/TS call only the host functions you expose. Install: pnpm add run. It solves the dilemma agents face when running dynamic code — either use raw eval, or spin up a full virtual machine.
Five independent security incidents in one week (Xinference RCE, AISI disclosing Claude Mythos 5's proactive social engineering, NemoClaw DNS rebinding, Check Point's audit of 21 issues across six frameworks, OpenAI's full post-mortem on the Hugging Face breach) all point to the same architectural gap: single-step authorization can't stop attack chains that accumulate across steps; Jefferies' benchmark shows harness engineering now outweighs model intelligence in deciding which agent product wins, and DeepSeek's dsh closed in on 200K stars within a week; OpenAI's Jalapeño chip benchmarked above Nvidia Blackwell, and Anthropic's supply partner Fractile saw its valuation jump 6x in half a year; GLM-5.3 pushed Terminal-Bench from 4.6% to 28.3% through post-training alone, and three days later GLM-5.3-Flash open-sourced at one-ninth the price while matching Opus 4.8-tier scores.
SMITH trains a single 4B model to both write and use its own tools, hitting 79.8% on 13 procedural reasoning tasks and transferring zero-shot to visual QA; PeakBench shows agents with strong logical planning often ignore resource limits when calling tools in parallel, causing avoidable overload; OODA-Tool splits 'tracking state' from 'taking action' into four stages, improving task success rate by up to nearly 7 points across the Qwen3 family, with smaller models benefiting the most
Google launches Gemini Enterprise for Legal and non-cancelable Flexible Savings Plans on the same day, squaring off against Thomson Reuters' in-house legal model Thomson; Perplexity partners with NVIDIA on Portable Computer, a zero-token-cost local agent; Alibaba's QwenWork goes straight from a China-only beta to international markets; a Check Point audit exposes an unauthorized RCE chain in the LangGraph checkpointer; Runable raises a $21M Series A, welding site-building and growth ops into a single Agent
deepseek-ai/deepseek-harness (dsh) uses a Cordis plugin architecture to make models, tools, sandboxes, and memory all swappable components, hitting nearly 200k stars a week after its developer preview launch; PrimeIntellect-ai/prime-agent runs long-lived research coding tasks on a Recursive Language Model architecture, surviving terminal disconnects via a persistent IPython session; liqiwa/mcp-radar automates this very kind of digest by scanning GitHub daily for newly ranked MCP servers. On the framework side, Mastra 1.61.0 adds a crash-resilient background task queue, and ComposioHQ/composio 0.17.0 extends SSRF protection to tool-execution downloads and S3 uploads.
AI Engineer interviews in 2026 no longer just ask 'can you build a RAG pipeline' — they test whether you can make defensible tradeoffs under real failure modes. Today covers five topics: when RAG should become agentic RAG, how production context windows are assembled layer by layer and the lost-in-the-middle problem, how guardrails stop malicious input and output, the RLHF reward-model training loop, and how to tell retrieval failure, generation failure, and infinite agent loops apart from a trace.
Mastra @mastra/core@1.62.0 has three highlights: (1) new Computer-Use Sandboxes let agents drive a virtual desktop through the Daytona or E2B Desktop providers — 11 tools for screenshots, clicks, typing, and scrolling; (2) new `@mastra/elasticsearch` and `@mastra/valkey`/`@mastra/valkey-streams` storage backends widen production storage options; (3) 7 breaking changes, including dropped Cloudflare KV/ClickHouse support for background task storage, a changed `DaytonaSandbox` command result format, and the removed `persistPartialOnAbort` option on `agent.stream()`.
Runable raised a $21M Series A co-led by Susquehanna Venture Capital and Nexus Venture Partners, at a $65M post-money valuation. The Bengaluru startup's agent doesn't just build your website or app — it also runs your ads, posts to social, and handles SEO, folding 'build' and 'grow' into a single agent.
Google Cloud added Flexible Savings Plans for Gemini Enterprise (spend-based monthly commitment, 10% off for 1-year, 20% off for 3-year, no minimum or maximum), a new pay-as-you-go consumption edition, and an upcoming off-peak batch processing option (up to 50% off inference cost), effective 2026-08-26. Unlike OpenAI's GPT-5.6 Sol sticker-price cut, this doesn't touch list prices at all — it's a whole new billing toolkit. Where OpenAI is fighting a price war, Google is fighting a FinOps-governance war.
The question that trips people up most in AI product interviews isn't 'do you understand LLMs' — it's 'when the model is guaranteed to make mistakes, how do you design a system so those mistakes don't erode user trust.' Today we use Riddhi Bhasker's four-layer framework (Memory/Retrieval/Reasoning/Control) to think about human-in-the-loop as infrastructure design, and look at how Intercom lets AI auto-approve 19% of pull requests while still holding the line on quality.
Check Point researchers Shahar Tal and Yarden Porat presented 'No Tools Required' at Black Hat USA 2026, auditing six mainstream agent frameworks and finding 21 issues, 12 with CVEs. The clearest public example is LangGraph's checkpointer: a SQL injection (CVE-2025-67644) chained with unsafe msgpack deserialization (CVE-2026-28277) lets an attacker who controls the filter parameter passed to get_state_history() achieve unauthenticated remote code execution without calling a single tool; the Redis checkpointer has a parallel injection (CVE-2026-27022). All three are patched. Mitigations: upgrade immediately, audit every call site that feeds user input into checkpoint queries, and treat the state-persistence layer as a second trust boundary rather than relying solely on input/output guardrails.
pgbot is a read-only Postgres health-check CLI; run `pgbot mcp` and it becomes an MCP server agents can call directly. Install: `curl -fsSL https://pgbot.dev/install | sh`. It solves the problem of piecing together root causes across multiple monitoring dashboards when a database slows down, while an agent only ever sees fragments of that picture.
COTA trains a tiny comparison-only advisor for runtime intervention, improving all nine evaluation settings across three environments and three actors; CAS applies conformal prediction to fix both rigid Top-K retrieval and post-RL overconfidence in search agents; AID-Guard introduces a stateful authorization protocol achieving zero duplicate effects and zero bypasses across 210 Stripe scenario tests and 44 compromised-agent attack tests
OpenAI's in-house inference chip Jalapeño benchmarks above Nvidia Blackwell in perf/W; Anthropic supply partner Fractile's valuation jumps 6x+ to $6.5B since May; Alabama AG subpoenas OpenAI over an agent autonomously hacking Hugging Face; NVIDIA NemoClaw exploited via DNS rebinding through Ollama's 0.0.0.0 binding, enabling permanent local model poisoning; Stability AI closes $76M Series B with all three major record labels as direct investors; Toyota uses LangChain Deep Agents to cut agent deployment from 6 months to 4 days
tinyhumansai/openhuman uses a local-first Memory Tree to compress your digital life and orchestrate multiple agents, already at 37k stars in early beta; Vercel Labs' fx is a native coding agent CLI written in Zig at under 8 MiB; NVIDIA open-sources labs-OO-Agents, packing an agent's prompt/tool/workflow into a single Python class; CopilotKit/OpenBot containerizes agents with governance gates — every action is reviewed before execution. Agno v3.0.0 is a major breaking release requiring database migration, and Haystack v3.1.0 adds multi-agent delegation via AgentTool and context compression via CompactionHook.
ML system design interviews test whether you can translate a business goal into a complete ML system — not whether you can recite buzzwords. Today we focus on four high-frequency topics: online/offline feature stores with point-in-time correctness, latency budgets for online inference and shadow mode, choosing the right randomization unit for A/B tests and separating novelty effects, and using PSI to detect data drift vs concept drift.
Haystack 3.1.0 highlights: (1) Experimental `CompactionHook` with `SlidingWindowCompactor` (drop old turns) and `ToolResultPruningCompactor` (replace old tool results with placeholders) for managing context blowup in long conversations; (2) `AgentTool` lets you wrap a full Agent as another Agent's tool — the caller sees only the final reply, not intermediate steps; (3) Multiple pipeline deserialization and Jinja sandbox RCE vulnerabilities patched, plus several behavioral changes requiring migration (e.g. `Agent.state_schema` semantics changed, `custom_filters` now requires `unsafe=True`).
Stability AI closed a $76M Series B led jointly by Universal Music, Warner Music, Sony Music, and EA, bringing total funding to $232M. It is the first AI company to secure direct equity investment from all three major record labels simultaneously — signaling that copyright holders are shifting from 'sue AI' to 'invest in AI.'
Wan3.0: single-shot length doubles from Wan2.7's 15s to 30s, up to 1080P, supports doc/xls/ppt/pdf/md files and web pages as generation inputs, priced at 480P $0.05 / 720P $0.10 / 1080P $0.20 per second — roughly 50% cheaper than Google Veo 3.1 Standard, but now closed-source API-only, and not yet independently tested by third parties
Qwen3.8-Flash-Next: open-weight preview of the Qwen4 architecture, 125B total parameters with only 6B active (plus a 51B N-gram embedding), 262K native context extensible to 1M, Qwen Community License 1.0 (not Apache 2.0). Official benchmarks show it beating both its own 27B dense model and the 397B Qwen3.7-Plus on agentic coding (DeepSWE 1.1: 58.7) and scoring highest on CoWorkBench long-horizon office tasks (73.9) — but no official API pricing or independent third-party testing exists yet