8 stats that show AI got smarter faster than it got safer

SWE-bench Verified moved from 60% to just about 100% in a single yr. 

Wow…

A benchmark constructed to separate frontier labs from everybody else stopped doing that job in roughly the time it takes most enterprises to resume a cloud contract.

Capability is the a part of AI that retains fixing itself.

Stat 1: The benchmark that stopped which means a lot

That SWE-bench leap comes from Stanford HAI’s AI Index Report, alongside a blunter discovering: business produced 91% of notable AI fashions within the tracked interval, up sharply from prior years. Capability positive aspects have turn into the simple a part of this story.

A near-perfect SWE-bench rating leaves open whether or not that identical mannequin can sustain a coherent multi-turn conversation, how usually it still gets things wrong once deployed, and whether or not anybody can really demonstrate that reliability past the benchmark itself.

The report’s personal framing lands more durable than most vendor decks handle: a discipline scaling faster than the programs round it can adapt. Somewhere, a advertising workforce is already turning that sentence right into a keynote slide, lacking the purpose on the way in which to the font alternative.

Stat 2: The quantity that ought to fear a CISO extra than the benchmark rating

Veracode’s GenAI Code Security Report examined extra than 100 fashions and located the common safety move fee sitting at 56%, primarily flat towards the prior yr. Java code failed safety exams over 70% of the time.

Models write code that compiles at near an ideal fee. Writing code that resists an attacker turned out to be a separate ability totally, one that issues most as soon as that code reaches mission-critical systems. The compiler has zero opinion in your firewall.

The identical disconnect exhibits up in agent evaluations, the place fashions keep failing simple selection traps whilst headline scores climb.

Stat 3: The transparency rating transferring in the other way from functionality

Stanford’s Foundation Model Transparency Index fell from 58 to 40 factors this yr, and 84% of essentially the most notable latest fashions shipped with the training code omitted. Google, Anthropic, and OpenAI have all stopped disclosing dataset sizes and coaching durations for his or her latest releases.

The labs topping the leaderboards are the identical labs publishing the least about how they got there.

Stat 4: The cash that says no person is ready for the security information

Two figures put a quantity on how a lot confidence the market is putting in functionality alone, forward of the belief query catching up.

• Global company AI funding hit $581.7 billion, up 130% yr over yr, with generative AI funding alone reaching $170.9 billion, in accordance with Stanford HAI.
• The prime 4 US hyperscalers pushed mixed information heart capex towards $600 billion, together with Meta, on a path towards $1.7 trillion globally by 2030, per Dell’Oro Group.

Increasingly, that spending follows a “train once, infer forever” logic that is reshaping how budgets get allotted.

Stat 5: The language shift that occurred faster than most roadmaps up to date

GitHub’s Octoverse report discovered TypeScript overtook each Python and JavaScript in August 2025 to turn into the platform’s most used language, the most important language shift GitHub has recorded in over a decade. Somewhere, a Python maintainer is recalculating a profession plan mid-standup.

A five-year language roadmap turned a two-year one, the type of shift AI architects now plan around faster than most individual engineer checklists get up to date.

Stat 6: The ability cluster that outran the org chart

Lightcast’s contribution to the Stanford AI Index tracked mentions of the “agentic AI” skill cluster in US job postings and located them rising over 280% in a single yr, transferring from 0.06% of postings to 0.23% (roughly 90,000 postings).

Hiring managers are asking for a ability set that barely had a reputation eighteen months in the past, whereas loads of resumes nonetheless listing “proficient in Excel” because the differentiator proper above the half the place they declare they will wrangle an agent swarm.

Stat 7: The workforce quantity sitting in survey information earlier than it hits headcount

The identical Stanford Index discovered a 3rd of surveyed organizations anticipate AI to shrink their workforce within the coming yr, concentrated in service operations, provide chain, and software program engineering.

Large-scale job losses have but to seem in combination employment information, a niche the report treats as a timing problem somewhat than a reassurance. That distinction issues extra than both half of the sentence learn alone, and it is value revisiting on the subsequent earnings season somewhat than this one.

Stat 8: The adoption line that crossed mainstream earlier than governance caught up

Organizational AI adoption reached 88%, up from 78%, in accordance with the same Stanford report. Almost each group value surveying has adopted one thing, which in loads of instances means a pilot license no person has opened because the kickoff assembly.

Adopting one thing differs meaningfully from capturing enterprise value from it, and differs additional nonetheless from constructing the agent experience layer most of those programs nonetheless want. That hole is strictly what exhibits up within the rising listing of agentic deployment mistakes enterprises hold repeating.

So, what ought to builders and patrons do with the hole?

• Treat benchmark scores as a single enter amongst a number of. SWE-bench close to 100% and a safety move fee caught at 56% describe the identical mannequin era. Procurement constructed on functionality scores alone is studying half the web page, a lesson each (*8*) learns early, normally proper after a vendor demo that skipped safety totally.
• Ask distributors what they stopped disclosing, alongside what they shipped. A transparency index falling from 58 to 40 factors is a sourcing query each technical purchaser must be asking out loud.
• Budget for the safety evaluate the Veracode quantity implies. A 56% move fee means roughly half of dedicated AI-generated code wants the scrutiny a workforce offers human-written code, at minimal.
• Hire for the ability cluster earlier than the job posting catches up. A 280% leap in agentic AI mentions is a number one indicator value appearing on earlier than it turns right into a expertise scarcity, particularly whereas the brokers themselves hold tripping over basic “why” questions. Interview accordingly.

Capability was at all times going to maneuver first. The genuinely arduous drawback, the one these numbers go away open, is closing the gap between how good these programs have turn into and the way a lot anybody really is aware of about them.

That drawback lands on engineers first, however it lands simply as arduous on the chief AI officers steering strategy from Silicon Valley and the tech leaders doing the same from New York. Progress got its keynote. Trust continues to be ready for its activate stage.

Where the infrastructure cash really has to land

The capex figures above describe the highest of the funnel. What occurs between a hyperscaler writing a test and a manufacturing AI system operating is a modernization drawback most infrastructure groups are nonetheless fixing in actual time.

AIAI’s report, Bridging the gap from supercomputing to AI factories, pulls direct perception from NVIDIA and WEKA management on that precise hole: the place HPC design assumptions break below manufacturing AI workloads, what modernization appears to be like like as soon as the test clears, and the way groups hold present infrastructure funding intact whereas making the shift.

Download the report earlier than the following infrastructure finances will get accepted on assumptions value checking first.