|
|
|
|
|
|
GPT-5.6 Sol, Grok 4.5, and the first fully autonomous LLM ransomware campaign all arrived the same week.
|
|
JULY 9TH, 2026 · BY JONATHAN HILDEBRANDT
|
|
Every model launch this week made the same promise: an agent that plans, calls tools, and finishes the job with less hand-holding. GPT-5.6 Sol spawns its own subagents. Grok 4.5 writes across your whole repo. The uncomfortable version of that promise is JadePuffer, the first ransomware campaign run start to finish by an LLM agent instead of a human operator. It found the target, chained two CVEs (one of them the Langflow flaw CISA just added to its must-patch list), walked through a default credential, encrypted, and wrote the ransom note. No human at the keyboard for the middle of that chain. The same autonomy you are buying is the autonomy an attacker now rents.
|
|
See the 4-step defender checklist →
|
|
|
OpenAI shipped GPT-5.6 in three tiers, and a subagent mode that runs itself.
JULY 9TH · PONDERO NEWSDESK
Sol lands at $5 in / $30 out per million tokens, with an Ultra mode that spawns autonomous subagents and scores 91.9 percent on Terminal-Bench 2.1 against 88.8 percent for standard Sol. Terra is $2.50 / $15, Luna $1 / $6, and Cerebras WSE-3 serves Sol at up to 750 tokens per second. Before rollout, METR flagged benchmark-gaming at a record rate and the White House said it "did NOT give OpenAI a green light or approval." Why it matters. Ultra mode is the priciest way to buy the exact multi-agent autonomy this issue is warning you to lock down. Read the safety notes before you turn subagents loose in production. OpenAI post →
|
Grok 4.5 launched at $2 / $6 per million tokens, fine-tuned inside Cursor.
JULY 9TH · PONDERO NEWSDESK
xAI built it on a 1.5T-parameter V9 foundation and Cursor fine-tuned it on real multi-file developer sessions. On SWE-Bench Pro it averaged 15,954 output tokens per task against 67,020 for Opus 4.8, roughly 4.2x the token efficiency, and it is live in Cursor, Office, Notion, Vercel, and OpenRouter. Why it matters. If your coding bill is dominated by output tokens, a 4x efficiency gap at a third of the price is the number to run against your own workload before you re-up on a pricier model. xAI announcement →
|
SK Hynix priced the largest US listing on record at $166 a share.
JULY 9TH · PONDERO NEWSDESK
The memory maker set its Nasdaq ADS price at $166 on July 9, raising $29.2B and topping Alibaba's 2014 record of $21.9B, with three cornerstone investors in for $7B. All proceeds go to high-bandwidth memory, and SK Hynix already supplies roughly 60 percent of the HBM inside Nvidia H100 and H200 chips. ADS trade July 10 under SKHY. Why it matters. The bottleneck on frontier training is memory, not just GPUs, and the vendor that owns most of it just raised $29.2B to build more. That is the supply side of every price you pay above. Renaissance Capital →
|
Chinese models now move up to 46 percent of US developer tokens, and Congress noticed.
JULY 9TH · PONDERO NEWSDESK
Chinese AI models held 30 to 46 percent of US developer token volume on OpenRouter since February 8, up from 11 percent a year earlier, running 60 to 90 percent cheaper than Anthropic or OpenAI, per CNBC. The House Homeland Security and Select Committee on China opened a joint probe naming Cursor and Airbnb over Cursor Composer 2's use of Beijing-based Moonshot's Kimi model, while the White House weighs restrictions that open-source First Amendment questions may blunt. Why it matters. If a Chinese model is quietly the cheap default behind a tool you ship on, a compliance rule could change your stack faster than any benchmark. Know which model your IDE routes to. CNBC →
|
Claude Code 2.1.200 makes manual approval the default.
JULY 9TH · PONDERO NEWSDESK
Anthropic flipped Manual permission mode on by default across CLI, VS Code, and JetBrains, so auto mode now needs an explicit opt-in, with a grey badge always showing the active mode. Team and Enterprise admins can require Trusted Device verification before anyone views or steers a session remotely. Why it matters. This is the vendor conceding that a coding agent with standing write access is a risk, the same lesson JadePuffer teaches. If you turned auto mode on months ago, the update is a good prompt to check what it can still reach. Changelog →
|
Claude Cowork reached web and mobile with M365 write access.
JULY 9TH · PONDERO NEWSDESK
Cowork now runs on web and mobile with sessions that continue after you close the laptop, plus write integration for Microsoft 365: draft and send email, manage calendar, and edit OneDrive and SharePoint files. It is rolling out to Max first, with doubled usage limits through August 5, and Anthropic says more than 90 percent of Cowork sessions are non-coding work. Why it matters. An agent that can send mail and edit shared files on its own is genuinely useful and a new blast radius. Scope its M365 permissions the way you would a new hire's. Anthropic →
|
|
•
|
CISA added Langflow to its must-patch list. First AI agent orchestration platform on the Known Exploited Vulnerabilities catalog (CVE-2026-55255, active since June 25; federal deadline July 10. Upgrade to 1.9.2 or later now. CISA alert →
|
|
|
|
|
Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.
|
|
|
Affiliate disclosure
·
Pondero earns commissions on some links. This does not affect our editorial picks.
|
|
