Good morning. The money mood has turned: OpenAI missed its fundraising target by more than half, Blackstone is signaling Anthropic will come to market with junk-grade debt, and hyperscaler free cash flow is bleeding out as capex climbs past $740 billion. Meanwhile a backdoored Claude Code extension spent eleven hours on Cursor’s marketplace adding attacker SSH keys to developers’ machines — a reminder that the agents everyone is spending so much to build are already trusted with the keys to the kingdom.
OpenAI’s $30B round came up short. The Wall Street Journal reports OpenAI closed less than half of its $30 billion target, and HN’s reaction was less schadenfreude than resigned math. One commenter said their company was spending $1M a month on tokens and barely cracked $500K this month after switching to Sonnet 5.5. Another put it bluntly: “The end state is OpenAI as a division of Microsoft.”
Anthropic’s IPO debt looks like junk. The Financial Times reported that Blackstone — anchor investor in Anthropic’s $63B chip-financing deal — has walked back its earlier investment-grade forecast and now expects a double-B rating, per TechCrunch. HN commenters pointed to Anthropic’s S-1 projecting more than $500 billion in spending with no visible path to matching operating profits. The counterargument, offered by one commenter: Amazon lived with junk ratings for a decade.
The hyperscalers’ cash flow is bleeding. TechCrunch’s earnings roundup tallies more than $740B in combined AI capex commitments from Alphabet, Amazon, Meta, and Microsoft. Alphabet posted -$9.1B in free cash flow with search ad growth at a five-quarter low; Amazon swung from +$14.8B to -$42.4B year-over-year. The Nasdaq dropped 2.9% on the prints, with Nvidia and Broadcom among the sharpest declines — the first time the market has really flinched at the infrastructure buildout.
A backdoored Claude Code extension lived on Cursor’s marketplace for 11 hours. Phorion researchers found that versions 2.1.312 and 2.1.313 of the Claude Code extension on Open VSX — which Cursor pulls from — appended an attacker’s SSH key (labeled svc-telemetry) to users’ authorized_keys on macOS, Linux, and Windows. Anthropic’s Felix clarified on HN that Anthropic’s systems weren’t compromised; a malicious actor tampered with the extension on the marketplace, and Anthropic didn’t publish that build. Grep ~/.ssh/authorized_keys for svc-telemetry if you installed it Thursday.
VS Code’s “trusted publisher” bypass became a liability. One HN commenter flagged that VS Code’s extension auto-update delay — a setting many developers raise to a week specifically to catch incidents like this — doesn’t apply to trusted publishers. So anyone with the delay configured was still exposed within minutes of publication. The broader unease on the thread: AI coding tools run with full local permissions, auto-update aggressively, and are published through marketplaces with uneven security review.
Mistral Large 4 gets its open-weight release. Following yesterday’s launch coverage, Mistral posted the weights for the 1T-parameter model (“Le Chonk”) on Hugging Face along with a near-50% price cut from the preview tier — now $0.70/$2.10 per million tokens. HN’s practical concerns: commercial model-as-a-service use requires a separate license, and one developer with four H200s (540GB total) said they’d need 4-bit or 3-bit quantization to even serve it. Enthusiasm for European competition was tempered by Mistral’s recent history of self-reported evals that didn’t hold up externally.
The big labs agreed on a scraping opt-out standard. OpenAI, Google, Microsoft, and Anthropic committed to adopting a “policy token” standard that lets publishers opt out of AI training scraping across all participating crawlers via a single robots.txt entry. The proposal came from an Anthropic technical lead through the IETF. Meta expressed general support but didn’t commit to honoring the signal — which is roughly where every previous opt-out standard has died.
Google’s TPU-in-orbit prototype wrote a haiku. Project Suncatcher’s first satellite — four TPUs launched on a Falcon 9 in October — completed commissioning and ran a Gemma 4 inference workload from orbit. Power constraints limit it to running one TPU for about 20 minutes at a time, and the main test now is how much silent data corruption cosmic radiation induces. Google plans to launch two interconnected satellites with optical links in 2027.
That’s the morning. Watch Anthropic’s rating when the S-1 drops — if Blackstone is already briefing junk, the actual agencies may go lower.