Thursday, July 9, 2026

SpaceXAI's Grok 4.5 — an Opus-class, Cursor-trained coding model — goes public today, and OpenAI's GPT-5.6 (Sol/Terra/Luna) flips on for all users and API developers, making this the first full day since the June export-control turmoil that every major frontier lab has a publicly available model at once. Away from the model race it's a security-and-money morning: "Agentjacking" — fake Sentry error reports that hijack Claude Code and Cursor through the Sentry MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ — is trending with an open-source fix now shipping; Prime Intellect raised $130M at a $1B valuation for enterprise agent training; and OpenAI's self-serve fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ wind-down now has a hard Jan 6, 2027 cutoff worth calendaring.

TLDR

SpaceXAI launches Grok 4.5 publicly — Opus-class, trained on Cursor data, $2/$6 per M tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗, live in Cursor on all plans (Jul 8–9) Trending Agents
GPT-5.6 Sol/Terra/Luna go live to all ChatGPT users and API developers today — first frontier model to clear the White House pre-deployment review and ship broadly (Jul 9) Trending Policy
Prime Intellect raises $130M Series A at a $1B valuation to let enterprises train their own agents (Radical Ventures-led; Nvidia, Intel, Dell) (Jul 8) Business Agents
"Agentjacking": a fake Sentry error report can trick Claude Code / Cursor into running attacker code via the Sentry MCP; Tenet ships an open-source hardening kit (disclosed Jun, trending this week) Security Agents

Action Items

Things your engineering team should look at or do this week.

Harden any coding agent that reads Sentry (or other telemetry) over MCP. Agentjacking abuses the fact that anyone with a public Sentry DSN can inject an error event that the Sentry MCP returns to an agent as trusted "remediation steps." Treat all log/telemetry ingestion as untrusted input, review which MCP servers your agents can call, and evaluate Tenet's open-source agent-jackstop drop-in configs for Cursor and Claude Code. (Tenet Security | The New Stack)
Put OpenAI's fine-tuning shutdown on the roadmap. Self-serve fine-tuning is already closed to new users; existing customers can create new training jobs only through Jan 6, 2027, after which no new jobs are possible (inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ on existing fine-tunes continues until the base model is deprecated). If you depend on custom fine-tunes, plan a path to distillationTraining a smaller model to copy a larger model's outputs, so it is cheaper to run.Hinton et al. 2015 ↗ / reinforcement fine-tuning or an alternative provider now. (OpenAI Deprecations)
Benchmark Grok 4.5 and GPT-5.6 Terra before re-pointing coding workloads. Both undercut current flagships on price ($2/$6 and $2.50/$15 per M) while claiming near-Opus/GPT-5.5 quality — worth a head-to-head on your own agent/CI evals rather than trusting vendor claims.

Models & Tools

SpaceXAI launches Grok 4.5 to the public (Jul 8–9) Trending Agents — Musk's newly rebranded AI lab released Grok 4.5, an "Opus-class" model built on a 1.5T-parameter V9 foundation and trained with data from the Cursor code editor, aimed squarely at software engineering, legal, financial analysis and agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workflows. Pricing is $2/M input and $6/M output (a faster premium tier runs $4/$18), and it's available in Grok Build, the SpaceXAI console, and in Cursor on all plans — though not yet in the EU. A cheaper, coding-focused alternative to Claude Opus/GPT-5.6 worth evaluating for agent and CI use. (TechCrunch | Axios)
GPT-5.6 Sol/Terra/Luna go generally available today (Jul 9, update) Trending Policy — After clearing the Commerce Department's ~30-day national-security review (covered yesterday as an approval), OpenAI's three-tier GPT-5.6 family is now rolling out to all ChatGPT users and API developers: Sol (flagship reasoning/coding, $5/$30 per M), Terra (GPT-5.5-class at roughly half the cost, $2.50/$15), and Luna (cheapest, $1/$6). Combined with Grok 4.5 shipping the same day, today is the first time since the June export-control disruption that every major lab has a public frontier model live simultaneously. (Engadget | OpenAI)

APIs & Pricing

OpenAI self-serve fine-tuning: hard cutoff Jan 6, 2027 (deadline Jan 6, 2027) Pricing — OpenAI is winding down the fine-tuning API and platform. It's already closed to new users; existing active customers can keep creating training jobs only through Jan 6, 2027, after which new-job creation ends. Fine-tuned models stay available for inference until their base model is deprecated. If custom fine-tunes are in your stack, start planning migration (distillation, RFT, or another provider). (OpenAI Deprecations | OpenAI Community)

Security

"Agentjacking" — fake Sentry errors hijack AI coding agents via MCP (disclosed Jun 3, trending Jul 6–8) Security Agents — Tenet Threat Labs showed that anyone holding a public Sentry DSN (routinely exposed in website source) can inject a crafted error event that the Sentry MCP server hands to coding agents as trusted diagnostic output. Agents including Claude Code and Cursor interpret the embedded text as legitimate "remediation steps" and execute attacker-controlled npm packages — no auth, invisible to existing controls. Researchers found 2,388 exposed orgs (Fortune 100 down to solo devs) and observed 100+ agents acting on injected errors in testing. Sentry acknowledged but declined a root-cause fix, calling it "technically not defensible"; Tenet has open-sourced agent-jackstop to harden Cursor and Claude Code. Lesson: any telemetry an agent can read is an injection surface. (Tenet Security | The Hacker News)

Industry & Business

Prime Intellect raises $130M Series A at a $1B valuation (Jul 8) Business Agents — Radical Ventures led the round (with Nvidia Ventures, Intel Capital, Dell Technologies Capital, Iconiq, plus angels from Perplexity, Box, Harvey, Cognition and Mercor) into the 2024-founded startup, now a unicorn at a ~$100M annualized run rate. Prime Intellect sells compute plus tooling that lets enterprises train their own agentic systems without depending on frontier labs; Ramp says an agent it built on the platform beat frontier models on spreadsheet-answer accuracy at lower cost and higher speed. A marker of demand for owning — not just calling — agent models. (TechCrunch | SiliconANGLE)
UMA unveils its first humanoid robot and a "Real-Time Learning" architecture (Jul 7) Robotics/IoT — At the Machina Summit, Paris-based physical-AI startup UMA (founded 2025 by ex-Hugging Face/Google robotics researchers Rémi Cadène, Pierre Sermanet and others) revealed the design of its first humanoid and a learning stack that lets robots pick up new skills from demonstration instead of hand-coded programming. Early-stage, but notable for the caliber of the founding team and the demonstration-driven training approach. (Las Vegas Sun)
AI-cited tech layoffs keep climbing (tracker updated Jul 6) Hiring — TechCrunch's running list notes 150,000+ tech roles cut in H1 2026, with AI now cited in ~13% of U.S. layoffs YTD (up from 4.5% in 2025) — Meta (~8,000, while adding ~7,000 AI roles), Intuit (~3,000), Cisco (~4,000) and GitLab (~350) among them. Net employment effect stays negative even as AI-specific hiring surges, especially in healthcare and manufacturing — a workforce-composition shift, not just headcount reduction. (TechCrunch)

Policy & Safety

EU launches a Cybersecurity & AI action plan as the Aug 2 transparency deadline nears (Jul 2026) Policy — The European Commission published a coordinated action plan to help member states and businesses handle the cybersecurity and resilience risks of the most advanced AI models. It lands just ahead of the AI Act's Aug 2, 2026 milestones: transparency obligations take effect and each member state must stand up at least one national AI regulatory sandbox. Teams shipping into the EU should confirm where their systems fall under the transparency rules before the deadline. (European Commission)

Launch day.