Anthropic published metrics showing Claude has crossed from tool to collaborator. As of May 2026, Claude wrote more than 80% of Anthropic's production code; the same engineers shipped eight times more code per quarter than the 2021–2025 baseline. The model now completes 76% of open-ended engineering tasks successfully, up 50 percentage points in six months. Doubling times on capability benchmarks have compressed to roughly four months. Anthropic frames three futures from here: capabilities plateau, efficiency compounds under human guidance, or AI begins designing its own successors without a stable stopping point. The report focuses on the third scenario's risks. When a system can improve the model it runs on, and that improved model improves the next generation, human oversight becomes harder to maintain at each step. Anthropic is not claiming that threshold has been crossed; it is arguing that the current trajectory makes a clear plan necessary before it is.
Recursive self-improvement has been a key concern in AI safety for decades: a system that can meaningfully upgrade its own successors could accelerate in ways human engineers cannot track. Anthropic's report is significant because it provides concrete productivity data rather than theoretical speculation, grounding the concern in what Claude is already doing inside the company.
A drug called daraxonrasib nearly doubled median survival for patients with advanced pancreatic cancer in a 500-person clinical trial, according to results published in the New England Journal of Medicine and presented at the American Society of Clinical Oncology conference in Chicago on May 31. Patients on daraxonrasib survived a median 13.2 months; those on standard chemotherapy survived 6.7 months. The drug was developed by Revolution Medicines in Redwood City, California, and targets the RAS protein family, a cluster of proteins that drives roughly a third of all human cancers. Researchers had considered RAS proteins undruggable for decades because their shape gave inhibitor molecules almost nowhere to bind. Daraxonrasib works by locking onto a pocket that opens only briefly on the active form of the protein. Researchers from MD Anderson Cancer Center led the trial. The results represent the first major breakthrough in pancreatic cancer treatment in more than a decade, and researchers believe the approach could open paths to treating other cancers whose driver proteins were previously considered off-limits.
Pancreatic cancer is among the most lethal cancers because it is usually caught late and resists most treatments. The five-year survival rate has been around 12%. RAS mutations appear in about 90% of pancreatic cases, making this protein family a long-sought target. Success here may unlock similar approaches for other cancers driven by MYC, p53, and beta-catenin.
The Justice Department opened civil rights investigations into 15 medical schools on Wednesday, widening a campaign that began in March with three schools and escalated in May with accusations against Yale and UCLA. The new probes target unnamed institutions that receive federal funding. The DOJ, led by Harmeet Dhillon as assistant attorney general for civil rights, argues that many medical schools prioritize racial demographics over professional qualifications in admissions. The legal basis is Title VI of the Civil Rights Act, applied through the Supreme Court's 2023 ruling in Students for Fair Admissions v. Harvard, which banned race-conscious admissions at colleges and universities. Medical schools receive federal funding and are thus subject to the same prohibition, the department argues. The AAUP has encouraged schools to resist the data requests, calling the investigations an overreach. No school has been named publicly, and no charges have been filed.
The Supreme Court's 2023 Students for Fair Admissions ruling barred race-conscious college admissions. The Trump DOJ is now extending that ruling to medical schools. Ohio State, Stanford, and UC San Diego were the first targets in March 2026; Yale and UCLA were accused last month of unlawfully advantaging Black and Hispanic applicants.
The Republican-controlled House passed a concurrent resolution Wednesday ordering Trump to withdraw US forces from hostilities against Iran, creating the first formal legal challenge to the conflict. The resolution used a mechanism from the 1973 War Powers Resolution that allows Congress to end a military engagement through a concurrent resolution passed by both chambers, rather than through legislation the president can veto. Constitutional lawyers dispute whether that mechanism remains valid after a 1983 Supreme Court decision that struck down a different type of legislative veto, and Trump's lawyers are expected to argue it does not. If the Senate passes a matching resolution, the legal showdown would move to federal courts. The administration called the House vote unpatriotic. The conflict began after Trump ordered strikes on Iran in late 2025; Congress never formally authorized it.
The 1973 War Powers Resolution requires the president to notify Congress within 48 hours of committing forces to hostilities and limits unauthorized engagements to 60 days without congressional approval. Presidents of both parties have disputed the law's constitutionality, and it has never been definitively tested in court. This vote could produce that test.
Anthropic published an update to Project Glasswing, an initiative using Claude to scan open-source software for security vulnerabilities, and released the reference implementation as an open-source tool on GitHub. The model found 23,019 vulnerabilities across more than 1,000 open-source projects, including 6,202 classified as high or critical severity. Of those, 1,596 have been disclosed to maintainers across 281 projects, and 97 have been patched. The automated pipeline runs seven stages: build, reconnaissance, finding, verification, deduplication, report, and patch. The system generates proof-of-concept crash code for confirmed bugs and drafts fixes for maintainers to review. Anthropic is also offering a managed version called Claude Security for organizations that want to run the pipeline without setting up the infrastructure. The current focus is memory safety vulnerabilities in C and C++ code, which account for a disproportionate share of critical bugs in systems software.
Memory safety bugs in C and C++ code, such as buffer overflows and use-after-free errors, are among the most exploited vulnerability classes in real-world attacks. The open-source ecosystem contains billions of lines of legacy C code that maintainers rarely have time to audit systematically. AI-assisted scanning changes the economics of finding these bugs before attackers do.
Humans spread across the planet in roughly 300,000 years. For comparable genetic change to produce that geographic range in a typical mammal, evolution would have needed about 88 million years. The gap, around 300-fold, is the subject of a new paper in the Proceedings of the National Academy of Sciences by Charles Perreault of Arizona State University. Perreault analyzed range maps for nearly 6,000 mammal species, compared how lineage age and body size relate to geographic spread, then measured humans against that baseline. The conclusion: genetic evolution cannot account for how quickly humans occupied every continent, built distinct local practices, and adapted tools to new environments. Cultural transmission, passing better techniques and accumulated knowledge across generations without waiting for genetic mutations, explains the shortfall. Humans remain less genetically diverse than a single wild chimpanzee population, yet occupy terrain equivalent to all other mammals combined.
This study adds quantitative weight to a long-running debate about what made humans ecologically dominant. Anatomically modern humans appeared around 300,000 years ago in Africa. The finding that culture drove expansion reinforces the view that symbolic communication and social learning, rather than biological changes, distinguish our species from other large primates.
S&P Dow Jones Indices announced Thursday it would not loosen its entry requirements after a public consultation, meaning SpaceX and other large newly public companies cannot join the S&P 500 ahead of schedule. S&P requires newly listed companies to trade for 12 months, post four consecutive quarters of GAAP profit, and maintain a public float of at least 10% of shares. SpaceX fails on all three counts: the company reported a $4.94 billion net loss in 2025, plans to put only about 5% of shares on the public market, and has not yet listed. The 12-month rule matters because roughly $13 trillion in passive index funds must buy a stock automatically when it joins the S&P 500; early entry would have forced that buying before SpaceX established a stable trading price. Nasdaq and FTSE Russell both revised their rules to allow faster entry, creating a split between major index providers on how to handle mega-cap IPOs.
An index fund tracks a benchmark by owning the same stocks in the same proportions. When a company joins the S&P 500, every fund tracking the index must buy shares automatically, regardless of valuation. That forced demand can push a newly public stock sharply higher. The 12-month rule exists so the stock can find its natural trading range before that wave hits.
The Meta Oversight Board, the independent review body Meta created to adjudicate content decisions, published a ruling finding that Meta has been banning user accounts without adequate justification. The ruling followed hundreds of public comments from users who said their accounts were removed without a clear policy violation. The board found the problem systemic enough to require structural fixes, not adjustments to individual cases. Platformer reporter Casey Newton covered the ruling, noting it adds pressure on Meta at a time when the company is already navigating changes to its content moderation approach. Account bans differ from content removals, which the board has ruled on before; this ruling targets the broader category of account-level enforcement, where users lose access to all their content and contacts at once. Meta committed to implementing the board's binding decisions when the body was established.
The Meta Oversight Board was created in 2020 with a $130 million endowment after years of criticism that Facebook and Instagram moderated content inconsistently. Meta committed to follow its binding rulings. Critics have argued the board mainly reviews cases Meta chooses to send rather than addressing systemic problems; this ruling targets systemic behavior directly.
A study of 4.5 million federal civil cases filed between 2005 and 2026 found that AI-generated documents rose from 1% of filings in 2023 to 18% in 2026, driven mostly by people representing themselves in court. The share of self-represented litigants rose from 11% in 2022 to 16.8% in 2025. Federal judges are divided on how to respond. Some find that AI-assisted filings are cleaner and easier to process than handwritten ones, even when the legal arguments remain weak. Courts are also split on privilege: a Michigan court ruled ChatGPT discussions qualify as protected work product, but a New York court ruled that documents generated with Claude lack attorney-client privilege because AI companies share user data. A lawsuit filed in March by Nippon Life Insurance alleges that OpenAI's ChatGPT practiced law without a license by helping a user reopen a settled case. New York has introduced legislation barring chatbots from posing as licensed attorneys.
Self-represented litigants, people appearing in court without a lawyer, are common in civil cases where stakes are too low to justify legal fees. Judges must give their filings some latitude. AI-generated documents now look professional enough that judges must scrutinize them for fabricated case citations, a persistent problem in current AI models.
A team at Swinburne University of Technology published two papers in May claiming to have detected a primordial black hole, a theoretical type of black hole that would have formed in the hot chaos of the early universe rather than from a dying star. The candidate object, named Phoebe, sits in the outer halo of the Milky Way about 60,000 light-years away, has roughly three times the mass of Earth's moon, and is physically smaller than a human hair. The team spotted it through gravitational microlensing: in 2019, a star in the Large Magellanic Cloud briefly brightened as something with strong gravity passed between Earth and the star. The event lasted about an hour. The researchers monitored approximately 10 million stars over five nights and generated one terabyte of data. If confirmed, Phoebe would be the first primordial black hole ever detected, and a candidate explanation for some fraction of dark matter. Most physicists favor alternative explanations, including a variable star or a free-floating planet.
Primordial black holes are hypothetical objects that could have formed in the first second after the Big Bang when the universe was dense enough to collapse into black holes without a massive star. They have been proposed as a component of dark matter, which outweighs visible matter roughly five to one. No primordial black hole has been confirmed.
