OpenAI trims its raise to $20B as a ChatGPT takeover bug and governments' sovereign AI bets dominate.
OpenAI dominated the news cycle on three fronts. It reduced its fundraising goal to as much as $20 billion with the deal still unsigned. Researchers disclosed a ChatGPT account-takeover chain that earned a $5K bounty. And Axios reports it is preparing a post-midterms campaign for mandatory "speed limits" on powerful AI, with data centers rebranded as "AI compute campuses." The company is asking for regulation while its flagship product's web attack surface is being picked apart and its capital raise shrinks.
OpenAI has lowered the amount it is seeking in its latest round to as much as $20 billion, per CNBC. Negotiations are ongoing and the deal has not closed.
A smaller target at the company setting the industry's capex pace is a data point for every lab and investor underwriting compute commitments. Until the round closes, the final size and valuation are the numbers to watch.
Parameter Labs found a sandbox escape in ChatGPT's HTML artifact previews, exploitable through the MathLive library, that led to full account takeover. Separate coverage reports OpenAI paid a $5K bounty for a share-link bug that ran attacker code in the logged-in session of anyone who opened the link, which appears to be the same flaw.
Rendering model-generated HTML and third-party libraries inside a logged-in origin is a hard boundary to defend, and every artifact or canvas feature from any lab carries the same risk. A $5K payout for account takeover on a product with this user base looks low.
SemiAnalysis released GDPx, an agentic capability benchmark spanning 11 domains, including chip design, cybersecurity, finance, and RSI. Anthropic's Opus 5.5 leads with a 12% score.
A benchmark where the best model scores 12% leaves room to measure progress, unlike saturated evals. The domain mix, especially chip design and cybersecurity, targets work with direct economic and security consequences.
Canada is investing CAD1.6 billion in Cohere to build domestic AI capability independent of US and Chinese providers.
A national government is now the anchor backer of a frontier-adjacent lab. Expect other middle powers to copy the model and to demand local control of weights and data.
South Korean accelerator maker FuriosaAI raised $620 million at a $2.9 billion valuation, roughly four times its prior mark. The money funds its challenge to Nvidia in AI inference and training hardware.
Investors are paying up for credible Nvidia alternatives. The raise lands alongside Hyperbolic's $30M for multi-neocloud compute rental and Iluvatar's claims of beating the H200 in select inference tests.
Axios reports OpenAI will run a policy and PR campaign after the midterms, advocating mandatory speed limits for powerful AI and rebranding data centers as "AI compute campuses." Separately, the White House says it can withhold the full draft of its voluntary AI testing framework, citing national security and commercial exemptions.
The leading lab is trying to shape the rules before they are written, while the administration's own framework stays opaque. Mandatory limits would hit competitors with fewer resources hardest.
Canada backs Cohere with CAD1.6B, South Korea budgets 21.3 trillion won (~$16B) for sovereign models, chips, and data centers, and India floats on-premises "model travels to the data" deployment.
ChatGPT's artifact sandbox was escaped, Northflank's breach led to abuse of customers' AI API keys, and Talos documents the Volta infostealer using LLMs to rank victims by value.
Apollo's economist sees "zero financial payoff" for most companies, Minecraft agent swarms collapsed into bureaucracy, and Dazzle, Duckbill, and Wajo use humans when their agents fail.
Talos found the Volta infostealer using an LLM to appraise victims by financial value, which is concrete evidence of criminals building LLMs into their operations.
"GPT-5.6 Sol beats Factorio" is a single blog post on peaceful mode, which removes combat, so it tests planning rather than survival.
Watch whether OpenAI closes its round at or below $20B and whether its speed-limit campaign names concrete compute or capability thresholds.