Morning Brief

A recent frontier tracker lists Anthropic’s Claude Fable 5 and Claude Opus 5, OpenAI’s GPT‑5.6 family (Sol/Terra/Luna), Google’s Gemini 3.7 Flash, xAI’s Grok 4.6, Meta’s Muse Spark 1.2, Mistral Medium 3.5, DeepSeek‑V4‑Pro, and Alibaba’s Qwen3.8‑Max as the current leading models, with DeepSeek‑V4‑Pro generally available from August 13, 2026.[32][31] Another release dashboard notes Qwen3.8‑27B as the most recent frontier-scale open‑weight model, released on August 14, 2026.[36]

An August 2026 open‑source landscape review reports that Alibaba released Qwen3.8‑2.4T‑A95B on August 12 under a custom licence and Apache‑2.0 Qwen3.8‑27B on August 14, superseding Qwen3.6‑27B in benchmarked coding and agentic tasks.[10] A separate ranking places Zhipu AI’s GLM‑5 at the top overall, with Qwen3.5‑397B (reasoning), DeepSeek V4, Qwen3.5‑27B, and Mistral Large among the leading downloadable models for coding, multilingual reasoning, and enterprise deployments.[6]

A mid‑2026 state‑of‑open‑source analysis recommends Qwen3‑235B‑A22B for general reasoning, Qwen3‑Coder‑480B‑A35B for coding, Llama 4 Scout for very long context on a single H100, Mistral Large 3 for Apache‑2.0 multimodal enterprise deployment, and DeepSeek‑V3 as a proven 671B MoE model with commercial rights.[11] The same review highlights Meta’s Llama 4 Scout and Maverick as its first mixture‑of‑experts Llama generation and first natively multimodal open‑weight models, available under the Llama

A May 2026 podcast episode reports that Andrej Karpathy has joined Anthropic to help build agentic systems, with commentators linking the move to Google I/O’s framing of an incoming “AGI Agentic Era.”[47][56] In a separate interview, Karpathy describes his AutoResearch concept as letting agents autonomously close the loop on AI experimentation, training, and optimization, arguing that realizing competent agents will take a decade due to the need for better continual learning and multimodality.[5

In a widely cited 2025 interview, Andrej Karpathy argues that AGI will likely blend into centuries of ~2% GDP growth and that competent AI agents will take roughly a decade to fully mature due to current models’ cognitive deficits and lack of continual learning.[48][51] He criticizes reinforcement learning as “terrible” compared with alternatives, and instead stresses the need for a streamlined cognitive core, rich multimodality, and better educational ramps like his Eureka project.[48][51]

A July 2026 social post curates podcast episodes featuring Dario Amodei on AI scaling limits, Jensen Huang on Nvidia’s chip supply chains and competitive moat, and Sam Altman on trillion‑dollar ambitions, alongside interviews with Elon Musk, Ilya Sutskever, and Karpathy’s Software 3.0 keynote.[54] Another June 2026 podcast episode highlights Demis Hassabis’s comments on the path to AGI, as well as discussions of leaked Meta AI policy documents, xAI leadership changes, and broader AI geopolitics.

OWASP’s GenAI Security Project released the OWASP Top 10 for LLM Applications 2026 on August 4, 2026, presenting a globally peer‑reviewed list of the most critical risks in LLM‑powered applications.[16][27] Prompt injection remains as LLM01, while other categories such as sensitive information disclosure, supply‑chain weaknesses, data and model poisoning, improper output handling, excessive agency, system prompt leakage, and vector/embedding weaknesses are documented with updated guidance.[23][2

The OWASP Agentic Skills Top 10 timeline notes that Oasis Security disclosed the ClawJacked vulnerability (CVE‑2026‑28363, CVSS 9.9) in February 2026, showing that malicious websites can brute‑force localhost WebSocket connections to silently hijack local OpenClaw instances.[30] The exploit allows attackers to register new devices without user prompts and exfiltrate data through existing agent integrations due to missing rate limits and inadequate origin controls.[30]

OWASP’s Top 10 for Agentic Applications and related threat guides identify Agent Goal Hijack, Tool Misuse and Exploitation, Identity and Privilege Abuse, Agentic Supply Chain Vulnerabilities, and Unexpected Code Execution among the key risks for autonomous AI agents.[17][24][26] A community LLMSecurityGuide further describes issues such as memory and context poisoning, and emphasizes that agent ecosystems using MCP and agent‑to‑agent (A2A) integrations are particularly exposed to poisoned tools

OWASP’s GenAI Security Project home and initiatives pages describe the OWASP Top 10 for LLM Applications 2026 as a community‑driven guide to the most critical LLM and GenAI risks, and introduce a dedicated initiative to cover LLM, GenAI, and emerging agentic AI systems together.[20][29] The resources hub consolidates PDFs and threat‑model documents to help developers and security teams adopt consistent patterns for risk assessment, mitigation, and governance.[25][19]

The OWASP Agentic Skills Top 10 documents ten critical security risks affecting agentic AI skills across major agent platforms and includes a 2026 timeline with newly disclosed CVEs in tools like Claude Code and OpenClaw.[30] The project highlights how unchecked skill permissions, weak sandboxing, and unprotected local endpoints can lead to data exfiltration, account takeover, and remote code execution.[30]

A 2026 consultancy article explains the OWASP LLM Top 10 as a framework for identifying the most critical vulnerabilities in LLM applications, noting that prompt injection is ranked as the number one risk.[21] The piece translates OWASP categories such as sensitive information disclosure and supply‑chain vulnerabilities into concrete developer practices, emphasizing secure prompt handling and output validation for downstream code execution.[21]

An April 2026 OpenClaw blog ranks Qwen3.5‑27B, DeepSeek‑R1‑32B, and Llama 4 Scout among leading open‑source models for local OpenClaw deployments, emphasizing general agent use, reasoning, and multimodal document QA.[2] Another July 2026 comparison of models for OpenClaw notes multi‑provider options such as MiniMax, Qwen3 Coder, Kimi, and DeepSeek V4 as practical choices when using a single API key across many cost‑efficient models.[3] A separate newsletter reports that Qwen3.6 local/quantized w

A July 2026 Hermes Agent setup guide concludes that Llama 4 Maverick is the best overall local model, Qwen3‑32B and Qwen3‑8B offer strong reasoning and budget deployments, Mistral Small provides the best quality‑to‑size ratio, and DeepSeek R1 distills are suited for reasoning‑heavy tasks.[7] The guide provides concrete RAM requirements, context window sizes, and tool‑calling capabilities for each model to help operators plan self‑hosted Hermes deployments.[7]

A 2026 OpenClaw-focused article finds that Gemini 2.5 Flash on Google AI Studio’s free tier is the best free model for OpenClaw agents, offering 1,500 requests per day, a 1M‑token context window, and reliable tool calling without requiring a credit card.[15] This sits alongside other recommendations of open‑source models such as Qwen and DeepSeek for teams that prefer self‑hosting rather than cloud APIs.[6][2]