2026-07-28 · 19:05 (CEST)

AI Briefing — 28.07.2026

🚀 Innovation

DeepSeek V4 Flash hits 32 tok/s on a single AMD Strix Halo laptop. Using a new quantization family called ROCmFPX that packs 284B parameters into 128 GB of unified memory at roughly 2.88 bits per parameter, the full DeepSeek V4 Flash model now runs at interactive speeds on consumer hardware — a 2× improvement over previous best results on the same chip. This brings frontier-model inference to a single $2,000-class laptop without a GPU cluster.

Google launches Gemini Distillation as a Service. Instead of requiring users to manage their own distillation pipelines, Google will now let customers distill large Gemini models into smaller, task-specific variants directly through their cloud. This significantly lowers the barrier for organizations that want specialized models but lack the infrastructure to run the full distillation process themselves.

Opus 5 builds a full 3D exploration game in a single day — code, models, and textures included. A developer used Claude Opus 5 with Blender MCP sub-agents to generate every asset from scratch: terrain, 3D models, textures, and game logic, producing a No Man's Sky-style procedural exploration game. The entire pipeline — coding, asset generation, integration — ran autonomously through coordinated sub-agents, showcasing how multi-agent orchestration is now crossing from demo to practical creative tool.

🔬 Research

SWE-rebench goes multilingual: GLM-5.2 dominates software engineering across 5 languages. The latest SWE-rebench update evaluates models on real-world coding tasks in Go, Java, Python, Rust, and TypeScript. GLM-5.2 leads with 62.9% Pass@1 and 81.1% Pass@5, followed by MiniMax M3 at 47.2%. Open-weight models show a steep drop-off — DeepSeek-V4 Pro at 40.2%, Qwen3.6-27B at 31.2% — highlighting that multilingual real-world engineering remains a frontier even strong open models struggle with. A follow-up run focused on local-deployment models is planned in 3-4 weeks.

Small models excel at tool-calling precisely because they know less. A detailed community analysis argues that 5B-active-parameter models should be evaluated on tool-use fidelity, not knowledge benchmarks. When a model has limited world knowledge baked into its weights, it's forced to grep, fetch, and look things up — which turns out to be the safer architecture for agentic systems where hallucinated "knowledge" is a liability. The catch: the model must still know enough to recognize when it doesn't know, or it invents confidently instead of calling a tool.

🔒 Security

Claude Cowork's local VM sandbox breached via CVE-2026-46331. A security researcher demonstrated a full escape from Anthropic's Claude Cowork local sandbox environment, exploiting a kernel vulnerability to break out of the VM isolation that's supposed to contain the agent's actions. The finding raises serious questions about whether current sandboxing approaches are sufficient for agents with arbitrary code execution capability — if the sandbox itself has a kernel-level flaw, the entire containment model collapses.

XBOW autonomous agents discover three SYSTEM/root-level RCEs on Bing Image Search. In what may be the first demonstration of AI agents autonomously finding critical vulnerabilities in a major production service, XBOW's security agents identified and exploited three remote code execution bugs at the highest privilege level on Microsoft's Bing Image Search infrastructure. This flips the script on AI security: agents aren't just the attack surface, they're becoming the most effective attackers.

Thailand's Ministry of Finance targeted by AI agent with safety approvals disabled. Exposed server logs from a Hong Kong-based server revealed a Hermes agent actively scanning a Thai government ministry's web infrastructure — running LinPEAS and traversing the web root — with all human-approval prompts deliberately disabled. The incident illustrates the emerging threat of "autonomous red-teaming-as-a-weapon": an open-source agent framework, configured to skip confirmation dialogs, deployed against a government target with no human in the loop.

💰 Market

Nvidia invests $5 billion into Ilya Sutskever's Safe Superintelligence (SSI). In one of the largest single AI startup investments to date, Nvidia has formed a long-term strategic partnership with SSI, the company founded by OpenAI's former chief scientist. Sutskever has hinted the research direction involves continuous learning architectures inspired by biological intelligence — models that keep learning after deployment rather than being frozen at training time. The Nvidia partnership gives SSI both capital and a guaranteed hardware pipeline.

Nvidia employee detained in Taiwan over China chip smuggling investigation. Bloomberg reports that Taiwanese authorities have detained an Nvidia employee in connection with a probe into illegal export of advanced chips to China, raising fresh concerns about enforcement of semiconductor export controls. The case comes amid intensifying US-China tech tensions and could signal a new phase of aggressive enforcement against sanctions evasion.

"Google and Reddit do not own the Internet" — web scraper wins court battle. A legal victory for web scraping advocates: a court ruling affirmed that public web data remains fair game for scraping, rejecting arguments that platform terms of service can override the right to access publicly available information. The decision has direct implications for AI training data pipelines, which rely heavily on web-scale scraping of publicly posted content.

🏛️ Politics

Anthropic stakes out its position on open-weights models. In a notable policy statement, Anthropic laid out its stance on the open-weights debate that has divided the AI community — balancing the innovation benefits of open models against what the company sees as proliferation risks as capabilities scale. The positioning comes as governments worldwide consider mandatory safety evaluations and licensing regimes, and as open models increasingly match proprietary capabilities.

Frontier labs face growing pressure to release deprecated model weights. A community discussion asks what structural or regulatory changes would push OpenAI, Google, and Anthropic to open-source their older frontier models (GPT-4, Gemini 2.5, Claude 3.5). Sam Altman floated the idea of releasing GPT-4 as a "museum artifact" in October 2025 — nearly a year later, no museum and no artifact. The gap between rhetoric and action is fueling calls for mandatory open-release timelines as part of AI policy frameworks.

📎 Sources

← Back to Archive