Morning Brief

Mungomash’s August 21, 2026 frontier snapshot reports OpenAI’s GPT‑5.6 Sol/Terra/Luna family in general availability, Anthropic’s Claude Opus 5 as its latest Opus‑tier release (July 24, 2026), and Google DeepMind’s Gemini 3.7 Flash as a newly stable GA fast model as of August 13, 2026.[19][26] The same snapshot lists xAI’s Grok 4.6 (August 12, 2026), Meta’s Muse Spark 1.2 (August 5, 2026), DeepSeek‑V4‑Pro (April 24, 2026), Mistral Medium 3.5 (April 28, 2026), and Alibaba’s Qwen3.8‑Max (August 3,

AI Release Tracker’s August 2026 update identifies Qwen3.8‑27B by Qwen (released August 14, 2026) as the most recent frontier‑scale open‑weight model, while GLM‑5.3 by Z.ai (released August 14, 2026) is called out as the latest tracked GLM‑series frontier model.[16][21] These follow earlier open‑weight entries like DeepSeek‑V4‑Pro and Mistral Medium 3.5 that are now widely used in coding and reasoning stacks.[18][19]

PromptQuorum’s July 2026 Ollama update ranks DeepSeek‑V4‑Pro (April 23, 2026) as an algorithmic‑coding specialist with 93.5% LiveCodeBench under an MIT license, and recommends Qwen3.6‑27B (April 16, 2026) as a top dense 27B model for coding that reaches 77.2% SWE‑bench and fits in 24 GB VRAM at Q4 precision.[30] The same guide highlights Kimi‑K2.7‑Code and Laguna‑XS‑2.1 as other agentic, long‑context coding options for local workflows.[30]

In an October 17, 2025 episode of the Dwarkesh Podcast, Andrej Karpathy describes AGI as “still a decade away,” citing current models’ cognitive deficits, lack of continual learning, and limited multimodality as key blockers.[42][32] He also criticizes reinforcement learning as a fragile optimization method compared to the broader systems work needed around data, evaluation, and architecture.[42]

In a March 20, 2026 No Priors episode, Andrej Karpathy discusses code agents and his AutoResearch concept, where agents autonomously run the full loop of AI research tasks: experimentation, training, and optimization.[38] He frames this as part of a “loopy era of AI” where agentic systems continuously refine models and tools with minimal human intervention.[38]

Episode 216 of The AI Show (SmarterX) notes that Andrej Karpathy announced he was joining Anthropic in May 2026, following his prior roles at OpenAI and Tesla.[37][35] The episode situates this move amid ongoing legal disputes involving OpenAI and broader shifts in how companies adopt AI for automation.[37]

The OWASP GenAI LLM Top 10 2026, released August 4, 2026, documents prompt injection, sensitive information disclosure, supply chain compromise, data and model poisoning, improper output handling, excessive agency, system prompt leakage, and vector and embedding weaknesses as the most critical LLM application risks.[12][1][9] The 2025–2026 updates explicitly add and expand System Prompt Leakage and Vector and Embedding Weaknesses to cover RAG‑specific attacks and real‑world exploits against hidd

The OWASP Top 10 for Agentic Applications (version 2026, announced December 2025) identifies Agent Goal Hijack, Tool Misuse & Exploitation, Identity & Privilege Abuse, Agentic Supply Chain Vulnerabilities, Unexpected Code Execution, and Memory & Context Poisoning as key risks for autonomous AI agents.[2][4][11][9] OWASP’s Agentic AI Threats and Mitigations guide further emphasizes that dynamic MCP and agent‑to‑agent ecosystems make it easy for attackers to poison runtime components or trick agen

The LLMSecurityGuide GitHub project summarizes OWASP’s LLM and agentic taxonomies, listing LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure, LLM03 Supply Chain, LLM04 Data and Model Poisoning, LLM05 Improper Output Handling, LLM06 Excessive Agency, LLM07 System Prompt Leakage, and LLM08 Vector and Embedding Weaknesses alongside ASI01–ASI06 agentic risks such as Agent Goal Hijack and Memory & Context Poisoning.[9] It tracks each vulnerability’s description, risk level, and status, p

The OWASP GenAI Security Project’s LLM and GenAI Top 10 initiative describes a community‑driven effort to catalogue the most critical risks impacting LLM, generative AI, and emerging agentic systems, spanning prompt injection, supply‑chain compromise, vector store attacks, and autonomous agent failures.[10][15] It positions the LLM Top 10 2026 and Agentic Applications Top 10 as complementary guides for securing API‑exposed models and web‑integrated agent stacks.[12][2][10]

OpenLayer’s July 2026 OWASP LLM Security Testing guide explains that the 2025 OWASP update added Unbounded Consumption, System Prompt Leakage, and Vector and Embedding Weaknesses to account for RAG and agentic architectures.[3] It stresses that agentic systems require session‑level security testing, because prompt injection can propagate across multiple tool calls, leading to code execution and data exposure if outputs are not validated.[3]

ElevateConsult’s March 20, 2026 overview of the OWASP LLM Top 10 states that prompt injection ranks as the number one critical vulnerability for LLM applications, ahead of other issues like sensitive information disclosure and supply‑chain compromise.[5] The piece urges AI developers to treat LLM prompts and outputs as untrusted input and to add robust validation layers around them.[5]

OWASP’s Agentic Skills Top 10 notes that on February 26, 2026, Oasis Security disclosed ClawJacked, a high‑severity vulnerability (CVE‑2026‑28363, CVSS 9.9) in the OpenClaw ecosystem, alongside earlier CVEs affecting Claude Code.[13] The document categorizes such weaknesses under agentic skill risks and illustrates how compromised tools or descriptors can let attackers hijack agent behavior or execute remote code.[13]

Haimaker’s July 2026 guide to models for OpenClaw lists GPT‑5.6, GPT‑5.5, and GPT‑5.4 Mini from OpenAI as reliable general coding and reasoning backends, alongside MiniMax M3, Qwen3 Coder, Kimi K2.7, and DeepSeek V4 Pro for low‑cost and specialized coding loads.[29] The guide emphasizes that OpenClaw can route to multiple providers under one API key, making multi‑model coding agents more practical.[29]

PromptQuorum’s July 2026 local‑LLM report highlights Laguna‑XS‑2.1 (July 2, 2026) as an agentic coding model with a 256K context window and SWE‑bench Verified 70.9%, and Kimi‑K2.7‑Code (June 2026) as a coding‑focused agentic derivative of Kimi‑K2.6.[30] Both are positioned as strong bases for long‑horizon coding agents and local “vibe coding” workflows on consumer hardware.[30]