AI Research + Security — March 26, 2026
GPT-5.4 Pro Solved a Math Problem Open Since 2019.
Then a Supply Chain Attack Hit the AI Stack.
The same week frontier AI cracked a 7-year-old Ramsey-style hypergraph problem, attackers compromised the LiteLLM security scanner to steal credentials from 95 million monthly downloads.
Sources: Epoch AI (First AI solution on FrontierMath); Naskrecki performance analysis; Checkmarx LiteLLM supply chain report; March 2026.
OpenAI‘s GPT-5.4 Pro produced the first verified AI solution to a problem on FrontierMath: Open Problems, Epoch AI’s benchmark of research questions that mathematicians have tried and failed to solve. The problem comes from a 2019 paper by set theorists Will Brian and Paul Larson. It is a Ramsey-style hypergraph question: construct hypergraphs that establish improved lower bounds on a sequence H(n) arising in the study of simultaneous convergence of families of infinite series. That one definition covers both descriptions that circulated in coverage, because the hypergraph construction is the tool and the H(n) lower bound is the target. No solution existed in any literature. Kevin Barreto and Liam Price first elicited the solution from GPT-5.4 Pro, and Epoch AI verified it. Three other frontier models from Anthropic, Google DeepMind, and OpenAI subsequently solved the same problem under similar scaffolding. Epoch notes that none of them had one-shotted it in earlier runs, so elicitation mattered.
FrontierMath scores jumped from roughly 5% under GPT-4 in 2024 to 50% under GPT-5.4 Pro in March 2026 on Tier 1-3 problems (undergraduate to postdoc level). On research-grade Tier 4 problems, GPT-5.4 Pro scores 38%. Since Christmas 2025, 15 open mathematical problems have moved from unsolved to solved, with 11 (73%) credited to AI involvement. The gap between “scores well on problems with known answers” and “generates novel reasoning on unsolved problems” is closing faster than most mathematicians expected.
Two Results, Not One
Two distinct results landed in March 2026, and most coverage merged them. The first is the open-problem solve described above. Epoch classifies that problem as “Moderately Interesting” in its difficulty taxonomy, and FrontierMath: Open Problems structures each entry across three levels: a warm-up with known constructions, a challenge with no known construction that resists brute force, and a full problem requiring a general method for all values of n. GPT-5.4 Pro’s solution eliminates an inefficiency in existing lower-bound constructions. The model reportedly combined Python for generating candidate constructions with Lean-style formal checking during its search.
The second result is different in kind. Mathematician Bartosz Naskrecki contributed a Tier 4 FrontierMath problem built on 20 years of his own expertise across Galois theory, algebraic geometry, and arithmetic. He had proved the result himself in an unpublished 13-page solution roughly eight years earlier, then deliberately hardened the problem before submission and pre-tested it against o4-mini-high to confirm models could not touch it. Epoch AI ran GPT-5.4 (xhigh reasoning mode) on his problem eleven independent times. Ten attempts failed. The eleventh succeeded. Total compute across all runs was estimated at 5 to 15 million reasoning tokens. Naskrecki published a formal analysis of all eleven runs, “Performance Analysis of Repeated LLM Attempts at a Research-Level Mathematics Problem,” and called the pattern “a striking illustration of the last-mile problem in AI mathematical reasoning.” Each failed attempt explored a slightly different approach without reaching the critical insight. His problem was not open in the strict sense, since he held a proof. It was a research-level problem engineered to be unsolvable by machines, and a machine solved it once in eleven tries.
The verification chain matters as much as the solve. A model-generated solution is a candidate, not a proof, until humans check it. The pipeline runs: model output, automated verifiers, review by Epoch AI researchers, and final confirmation by the problem’s authors. Both March results cleared that chain. The reason the chain exists: GPT-5.4 Pro also appeared to crack a different Tier 4 problem, but preliminary investigation suggested it had located a 2011 preprint the problem author was unaware of. That is retrieval, not generation, and every claimed solve has to be screened for it.
The FrontierMath Benchmark and What It Measures
FrontierMath is a benchmark created by Epoch AI specifically to test mathematical reasoning beyond textbook problems. It contains 290 Tier 1-3 problems (with solutions for 237), 48 Tier 4 research-grade problems (solutions for 28), and a separate set of genuinely unsolved open problems. The key distinction: FrontierMath problems are designed so that the answer can be verified computationally but the solution path requires novel reasoning. This prevents models from pattern-matching against training data.
One structural issue deserves attention. FrontierMath was funded by OpenAI, which has exclusive access to all 290 Tier 1-3 problems and solutions to 237 of them, plus 28 of the 48 Tier 4 problems and their solutions. Epoch AI holds out the remainder. Whether OpenAI’s models were trained on any FrontierMath data is not publicly verified. This is a conflict of interest that should be factored into any interpretation of the benchmark results. The strongest evidence against contamination so far: on Tier 4, GPT-5.4 Pro solved 55% of the held-out problems OpenAI cannot access versus 25% of the OpenAI-accessible set, though Epoch notes neither difference reaches statistical significance at these sample sizes.
The Earlier GPT-5 Math Controversy
Context matters. In early 2026, OpenAI VP Kevin Weil posted (and later deleted) a claim that GPT-5 had “found solutions to 10 previously unsolved Erdos problems.” Mathematician Thomas Bloom, who maintains the authoritative Erdos Problems website, explained that the problems were not “unsolved” in the traditional sense. They were problems where Bloom was personally unaware of a paper containing the solution. GPT-5 had retrieved existing solutions from the literature, not generated new ones. Meta’s Yann LeCun called the situation embarrassing. Google DeepMind’s Demis Hassabis agreed.
The GPT-5.4 Pro result on the Brian-Larson problem is categorically different. The problem was genuinely open: no known solution existed in any literature, and Epoch’s verifiers and researchers confirmed the construction, with final sign-off resting on the problem’s authors. This is not retrieval. It is generation of novel mathematical reasoning. The distinction between “found an existing answer” and “generated a new answer” is the difference between a search engine and a reasoning engine. GPT-5.4 Pro demonstrated the latter on at least one problem.
What This Does and Does Not Mean
The Trajectory That Matters
From 5% on FrontierMath (GPT-4, 2024) to 50% (GPT-5.4 Pro, March 2026). A tenfold improvement in under two years. If this trajectory continues, frontier models will score above 80% on Tier 1-3 FrontierMath problems by mid-2027. The research-grade Tier 4 problems (currently at 38%) are the more meaningful measure. Those problems require generating reasoning that does not exist in any training corpus.
Since Christmas 2025, 15 open mathematical problems have been solved with AI involvement. That pace, approximately one every six days, suggests we are entering an era where AI-assisted mathematics becomes a standard research methodology. The analogy is computational proof assistants (Lean, Coq), which changed mathematics not by replacing mathematicians but by enabling them to verify longer and more complex proofs. AI is doing something different: it is generating candidate proofs and solution strategies that humans then verify and refine. The combination of AI generation and human verification may prove more productive than either alone.
Sources: Epoch AI, “First AI solution on FrontierMath” and FrontierMath: Open Problems documentation; Will Brian and Paul Larson (2019 paper posing the H(n) problem); Naskrecki, “Performance Analysis of Repeated LLM Attempts at a Research-Level Mathematics Problem”; OpenAI “Advancing Science and Math” case study; Winbuzzer GPT-5.4 Pro analysis; Garry’s List deep-dive; Computerworld; TechBuzz (GPT-5 Erdos controversy); 36kr (Terence Tao experiments). Updated 2026-08-18: An earlier version conflated two March 2026 results, attributing the Ramsey-style hypergraph open problem to Naskrecki and the H(n) definition to his Tier 4 problem interchangeably. The open problem is Brian and Larson’s 2019 H(n) lower-bound question, first solved by GPT-5.4 Pro with elicitation by Barreto and Price. The eleven-attempt story belongs to Naskrecki’s separate Tier 4 problem, for which the author held an unpublished proof. The article now defines each problem once, exactly, and describes the candidate-to-verified-proof chain explicitly.
The most telling detail in the entire episode is Naskrecki’s reaction. He is a domain expert with 20 years of specific expertise who publicly staked his reputation on AI being incapable of deep mathematical reasoning, calling it a very advanced calculator as recently as mid-2025. After watching GPT-5.4 work through his problem across 11 attempts, he reversed his position completely. Not partially. Not with caveats. Completely. When a domain expert who bet against AI changes their mind after examining the evidence, that signal is worth more than any benchmark score. The model did not just solve his problem. It changed his model of what AI can do.