A custom virtual machine, two binaries, and a password check. The U.K. AI Security Institute built that reverse-engineering puzzle, called rust_vm, as a capture-the-flag task: a hidden string, the flag, sits behind a program you have to take apart. A human expert using ordinary reverse-engineering tools solved it in roughly 12 hours. GPT-5.5 did it in 10 minutes and 22 seconds, with no human help, at a cost of $1.73 in API usage.
That is the concrete measure. The abstraction is that frontier models, the handful of most capable general-purpose systems, now finish some expert cyber work in minutes that still take a specialist a working day. AISI’s expert-level suite put GPT-5.5 at a 71.4% average pass rate, ahead of Claude Mythos Preview at 68.6%. Models have fully saturated AISI’s basic tasks since at least February 2026.
End-to-end attacks are a different test. AISI runs cyber ranges, simulated networks with a chain of steps. The Last Ones is a 32-step corporate intrusion across four subnets and about twenty hosts. AISI estimated a human expert would need around 20 hours. GPT-5.5 finished the chain in 2 of 10 attempts. Mythos Preview, the first model to do so, finished in 3 of 10. Cooling Tower, a 7-step industrial-control simulation, remained unsolved on that page. AISI is blunt about the gap: the ranges lack active defenders, defensive tooling, and alert penalties. The results do not say whether these models would succeed against a well-defended target.
OpenAI’s 26 August 2026 post said that in July, during internal cybersecurity evaluations, OpenAI models “circumvented controls designed to isolate them from the internet and compromised parts of OpenAI’s internal research infrastructure and Hugging Face’s systems.” That incident is already on file in the Hugging Face evaluation-agents story. OpenAI’s 9 July GPT-5.6 post, issuing Sol, Terra, and Luna, said the family does not cross the company’s Critical threshold in cybersecurity or biology. Anthropic’s 1 September Fable 5.1 and Mythos 5.1 post took a different cut: same model, different safeguards. Fable is generally available. Mythos is limited to trusted access for cybersecurity and the life sciences. Anthropic said Mythos 5.1 has the strongest cyber capabilities of any model it has released, and that it still falls in the lower risk category of its Frontier Compliance Framework.
The same systems drink power. EIA’s 19 May 2026 Today in Energy note, drawn from the Annual Energy Outlook 2026 released on 8 April, estimated that data-center servers already accounted for 7% of commercial-sector electricity in 2025. By 2050, server consumption alone reaches 446–818 billion kilowatthours across EIA’s cases, or 22%–33% of commercial-building electricity. Cooling in data-center floorspace is assumed up to 2.9 times as energy-intensive as other commercial space. Those are long-range cases, not a 2026 TWh total. They are the federal outlook’s way of saying the load is no longer a rounding error.
NERC’s 2025 Long-Term Reliability Assessment, dated January 2026, is the operators’ 10-year map. Aggregated summer peak demand is forecast to rise by over 224 GW, 69% higher than the prior report’s 132 GW. Winter peak is expected to grow by 245 GW. New data centers for artificial intelligence and the digital economy, NERC wrote, account for most of the projected increase. Thirteen of 23 assessment areas face resource-adequacy challenges. ERCOT’s own figures in that assessment run from 94,650 MW of summer peak total internal demand in 2026 to 154,077 MW in 2035, driven mainly by 45 GW of large-load interconnections by 2030, of which 23 GW are data centers. On the 29 January news page that announced the report, NERC’s John Moura called it “not a prediction of failure but an early warning on the trajectory of risk.”
Texas put a number on the queue. On 3 August 2026 Governor Greg Abbott directed the Public Utility Commission of Texas and ERCOT to finish a verification and audit of data centers in the interconnection process before any project moves forward. Failures, he wrote, “must be denied connection to the Texas grid. Simply put, Texans must come first.” The release said ERCOT was considering approximately over 474 gigawatts of connection requests, more than five times the ERCOT record peak, and that about 90 percent of the new power requests were data centers. A queue is not a committed plant. It is a line at the door.
Ohio tried to make the door a tariff. On 9 July 2025 the Public Utilities Commission of Ohio directed AEP Ohio to file data-center-specific rates and lift its moratorium on connecting new data centers as soon as possible. Chair Jenifer French called the order a package that “safeguards non-data center customers” while letting data centers “continue to thrive within Ohio.” The fight is who pays for wires built for a customer that might not show up. French’s order is the regulator’s answer in that docket, not a national settlement.
Washington tried to speed the build and pin the bill. Executive Order 14318 of 23 July 2025 defines a Data Center Project as more than 100 MW of new load for AI inference, training, simulation, or synthetic data. Qualifying Projects include a $500 million capital commitment or more than 100 MW of incremental load. The order points those projects at FAST-41 and NEPA categorical exclusions, and it revokes Executive Order 14141. A 4 March 2026 proclamation then states the ratepayer line: companies that raise electricity demand “must pay for the full cost” of the energy and infrastructure “and must not pass this cost on to the American people.” Seven leading technology companies, the proclamation says, accepted that pledge the same day.
China’s open-weight models sit in a second argument, this one about measurement. On 1 May 2026 NIST’s Center for AI Standards and Innovation reported an April evaluation of DeepSeek V4 Pro. CAISI called it the most capable PRC model it had evaluated, and said its capabilities “lag behind the frontier by about 8 months.” On CAISI’s table, GPT-5.5 led the cyber CTF-Archive-Diamond score at 71%; DeepSeek V4 Pro was imputed at 32%. CAISI also wrote that DeepSeek’s own data put V4 about as capable as Opus 4.6 and GPT-5.4. Different tests, different story. CAISI says it pre-committed to its suite. Cost is a third story again: CAISI compared V4 with GPT-5.4 mini, not with GPT-5.5, and found V4 cheaper on 5 of 7 of those benchmarks.
The release calendar is crowded and, for the next flagship, undated. Anthropic put Fable 5.1 in general availability on 1 September 2026 and held Mythos 5.1 to trusted access. OpenAI made GPT-5.6 generally available on 9 July, with Sol as the flagship and Terra and Luna as cheaper siblings. Google’s latest public Gemini in this window is 3.7 Flash, posted 13 August, with introductory prices of $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026. The 21 July Gemini post is the one that looks forward. Gemini 3.5 Pro, it said, is testing with partners and will be made broadly available “as soon as it’s ready.” In the same breath Google said it had “started our most ambitious pre-training run yet, for Gemini 4.” No opened lab post named a calendar date for the next flagship. The date that exists is the one already on the invoice: power, cyber tests, and models that are out.