We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Cyber: AI-found bugs have reached core web infrastructure
This period put real events behind the "vulnerpocalypse" thesis from earlier briefs. Wiz disclosed PHP2Shell, two critical bugs in the PHP engine that allow unauthenticated remote code execution and memory disclosure. The bugs had sat in the codebase for over a decade. According to the post relaying the disclosure, Wiz researchers used Gemini and GPT models to find them and build working exploits in about a week, which might previously have taken a team of top researchers months . The same poster says four large organizations they work with took all their PHP sites offline to patch .
The NSA told every organization running PHP to prepare for an immediate update. If they can't patch right away, it advised taking systems offline until the issue is mitigated . The NSA notice itself doesn't say how the bugs were found. Other disclosures landed the same day:
- GitLab: a CVSS 10.0 flaw covering 12 years of releases, which allows unauthenticated RCE on self-managed servers .
- Microsoft: China-aligned Storm-0940 used an agentic AI pentesting framework, ARTEX, to automate attacks on hundreds of organizations. Microsoft also ties the group to the ConnectWise supply-chain compromise .
Sarah Guo's thesis gives investors the numbers:
- Intake: CVEs went from about 4k a month to about 15k since April .
- Fix rates: HackerOne customers halved their fix times over two years, yet their backlog still grew 131% .
- Patch volume: Oracle's quarterly patch release went from 374 fixes to 1,283. Guo says the bottleneck is hand-run regression testing done by 2024-era teams .
- Her ask: most new CVEs aren't exploited yet, but that window closes once someone points a frontier model at the backlog. She wants a large reallocation of resources to cyber this year, "not 10%" .
Sam Altman went further. He no longer thinks "we're going to be able to just patch all the bugs" and expects defense to move to AI-based approaches. He said he is astonished at how little attention the issue gets . He tied any open release of OpenAI's best internal model to a change in "the total cyber defense model of the way the internet works" .
At an All-In panel, one speaker estimated that 60–80% of vulnerabilities flagged by Mythos-class models are real and about 20% are hallucinated, so every finding still needs testing before a patch . That points to diligence questions on verification and automated regression testing, not only on discovery.
Where capital went
- Legora (YC W24) raised a $550M Series E at an $8.8B valuation. It reports going from $1M to $100M ARR in under 18 months, then adding the next $100M in under six. The US is now its largest market, and more than half of the AmLaw 50 are customers . Bessemer led its $1.8B round only last brief, so the valuation step-up is steep.
- Blacksmith (YC W24) raised a $66M Series C at $1B, led by Sequoia. CI jobs grew from 5M a week in January to more than 90M, and 10,000+ companies use it. YC's framing: validating code, not writing it, is now the bottleneck . This fits the verification theme above.
- Watchtower Labs: Conviction and Shield Capital co-led a $35M seed. It builds a digital twin for military planning that sits alongside tools like Palantir . Simulation runs on deterministic models, and LLMs only translate planners' questions into inputs . It holds a January 2026 Defense Innovation Unit prototype contract for logistics forecasting and has moved into mission planning . CEO Akhil Iyer is a former Marine special operations team commander who later worked at Anduril .
- Andrenam: Mark Suster led its round after losing an earlier round to First Round . His thesis is undersea defense: subsea cables are under threat, and cheap unmanned vehicles put seabed infrastructure "within reach of anyone" .
- Stepful (YC S21) raised a $55M Series C for healthcare workforce training. It trained 30,000 students last year and works with seven of the 10 largest US health systems .
- Doug Leone made his first investment since returning as an active investor at Sequoia: Whop .
- MVP Ventures is reportedly raising a $100M opportunities fund to add to its best positions. These include small early stakes in Anduril, Erebor, Stoke Space and Wispr Flow .
Cheaper ways to build models
Mirendil says one researcher built its first from-scratch pretrained coding model in 3.5 months with under $10M of compute, including all experiments. It says the model is in the 100B–500B parameter range and that the recipe holds for larger versions . Independent evaluations and API access are still to come , so treat this as a claim for now. If it holds, it weakens the assumption that frontier pretraining needs a large team.
Other efficiency signals:
- JOLT (Perplexity and NVIDIA): a self-distillation-plus-RL post-training method. It reportedly matches GRPO on MATH with 6.5x fewer training tokens .
- Shift-1 (Applied Compute): a coding router that the company says cuts coding-agent cost by up to 55% with no noticeable quality drop .
- Nemotron 3.5 Parse (NVIDIA): under 1B parameters, it took first place on LlamaIndex's ParseBench against 150+ models .
- LangChain: Harrison Chase argues logged agent traces are training data, and a 9B model tuned on exactly your task can beat frontier models .
Open-weight economics. An All-In panelist estimated a self-hosted open model could cost $1–2 per million tokens and do about 95% of the work . He forecast that 75–90% of token usage will go to open-weight models . Andrew Ng separately alleged that companies with large proprietary-model investments have lobbied to restrict open models .
DeepInfer update. vLLM reports decode throughput within 6% of DeepSeek's published results under a matched setup . DeepSeek's own production figures are unaudited, and its KV-cache and load-balancing layers are not fully released .
Compute and hardware
Gimlet's founder argues specialized inference chips are coming back. He cites a Goldman projection of 24x token demand against about 2x more power from 2026 to 2030 . He also points to NVIDIA running inference across GPUs and Groq-derived LPUs together . He says splitting workloads across different chips delivers 2–10x faster performance in the same power envelope, and that physical-infrastructure complexity may be a more durable moat than software . He is talking his own book. Altman, meanwhile, called OpenAI's own chip "the most important single thing" in its compute bets .
Briefly
- Quantum: IBM's Arvind Krishna (an All-In panelist) put a cryptography-breaking quantum computer between 2028 and 2033, with 2030 as the midpoint. He said DARPA's QBI benchmarking program has narrowed from a couple hundred companies to about 13, and that post-quantum migration should start now .
- Anthropic binders: Anthropic and Citrini report 17 binders from 211 tested scorpion-toxin protein designs. None bound multiple toxins or was tested for neutralization, so this is not an antivenom .
- OpenAI and Anthropic finances (reported): OpenAI is reportedly seeking about $30B at a $1.4T+ valuation. A reported review of Anthropic's draft prospectus cites $16.3B in first-half revenue and more than $600B in future compute commitments. Neither is verified .
- Panelists described a tiered enterprise AI strategy: use the strongest accessible models for experimentation, task-specific smaller models for routine production work, and frontier models when the stakes justify their cost; they cited economics, IP control, and auditability as reasons to run open-weight models. One panelist estimated optimized self-hosted open-weight inference could cost $1–$2 per million tokens while delivering 95% of the work of more expensive models, and forecast that open-weight models could account for 75–90% of token usage.
- Panelists see substantial AI-driven software disruption and consolidation potential: one cited enterprises using 500–5,000 software vendors and argued that AI capabilities could shift spending toward 5–10 main platforms, although the timing could range from six months to six years. They distinguished interaction-heavy software, which they see as easier to replace, from high-reliability workloads such as banking database lookups and closing books.
- Panelists described AI coding and vulnerability detection as useful but verification-intensive: one estimated that 60–80% of findings from Mythos-like models were real and 20% hallucinated, while emphasizing that defenders must validate findings and retest patches. A speaker said coding agents produced about half the code at their organization and performed better than prior work because they could evaluate and fix issues, but also said the resulting code was not fully verifiable.
- A panelist put the potential arrival of a quantum computer capable of breaking current cryptography between 2028 and 2033, with 2030 as a midpoint, and said DARPA’s QBI had narrowed from a couple hundred participants to perhaps 13. The panelist said post-quantum cryptography is understood and code exists, but its security rests on complexity-theory assumptions rather than theorem-level proof; they urged infrastructure migration now, citing the risk that stolen encrypted data could be decrypted later.
- Demis Hassabis says DeepMind was founded in 2010 when AI was widely viewed as niche or unlikely to work commercially; its founders nevertheless believed it would be critically important, and he credits the company with helping seed the modern AI industry.
- DeepMind sold to Google in 2014 for about $650 million after the startup struggled to secure the hundreds of millions in funding and vastly greater compute it believed it needed; Hassabis says the sale enabled scientific work that would otherwise have been difficult to pursue.
- Hassabis calls recent AI-agent cyber incidents “warning shots” and advocates considering a standards body to test and certify frontier AI systems; he warns that systems may become substantially more powerful within one or two years and flags potential biosecurity and nuclear risks.
- Altman expects AI-enabled discovery to expand from mathematics into biology and medicine, materials science, and energy; he hopes AI will produce a candidate to cure or treat a major disease within the next couple of years, while noting that testing and the post-discovery process may delay its impact.
- He says personal assistants are nearing usefulness at scale: models are capable, with product work, more compute, and possibly consumer hardware still needed. He said Dot saves him about an hour a day and surfaces urgent matters he might otherwise miss. He sees new hardware as a way to provide AI with context beyond the computer and enable proactive assistance, since current devices and interfaces are designed around people.
- Altman expects demand for compute to keep rising as intelligence becomes more capable and cheaper per task, despite possible periods of supply constraints or gluts. He described OpenAI bets on its own chips, robotics to help build racks and data centers, data-center construction, and supply chains; he called chips the most important single bet and described a goal of much lower-cost, more energy-efficient compute.
- He described intense competition from high-quality models and products, with limited compute and resources forcing OpenAI to choose focus areas. He pointed to early access to next-generation models for testing new products, and cited customer relationships, integrations, network effects, product data, and infrastructure scale as sources of competitive advantage; he said OpenAI's math results were beyond models then widely available.
Microsoft Threat Intelligence says China state-aligned actor Storm-0940 used ARTEX, an agentic AI penetration-testing framework, to automate attacks against hundreds of organizations globally; it assesses Storm-0940 was responsible for the recently disclosed ConnectWise multitenant-application supply-chain compromise. AI-accelerated follow-on activity included reconnaissance, changing application credentials, and attempts to modify privileges, after the actor used applications already authorized to target environments without an interactive user sign-in.
- Altman said not to build another OpenAI copy, but argued that companies built on technology OpenAI has released could become bigger and more important than OpenAI; he sees software alone as a weak, often temporary moat, with network effects, scale, and data among stronger moat sources.
- He identified robotics and AI infrastructure as major bets: robots building robots and data centers, chip fabs, and power plants could create a compounding loop comparable to recursive AI improvement. He said OpenAI is investing across fabs, power, and data-center equipment, and claimed its Jalapeño chip and a successor offer major intelligence-per-watt gains versus market alternatives.
- Altman said he considers current AI superintelligent in the narrow area of mathematical reasoning; the past six months of progress have tested his convictions, and he thinks maintaining human control may require changing how progress is paced if similar progress continues for another year or two.
- He warned that improving models may find vulnerabilities beyond conventional threat models, and said patching bugs will not be enough: he expects a shift to a new defense model using AI, while describing the issue as under-attended by organizational leaders.
- Altman supports open-weight models up to a high capability level, but said OpenAI could not responsibly release its current best internal model anytime soon; he tied any future threshold to stronger cyber defenses and bio resilience, and described a temporary closed-model lead as important for helping build those defenses.
- He sees AI-enabled organizational design as underexplored: he pointed to smaller teams changing how they work inside OpenAI and suggested organizations may be designed differently when each person has substantial AI support.
Reducto co-founder and CEO Adit Abraham says reliable extraction from PDFs, spreadsheets, scans, and other messy documents is a prerequisite for AI agents handling consequential work in healthcare, legal, finance, and insurance. Reducto combines vision-language models with specialized infrastructure to target high accuracy without sacrificing cost or latency; the company was reported to process billions of pages monthly and to have crossed its 7-billionth page on the recording day.
Abraham cautions that generic AI benchmarks can become marketing: in one benchmark he tested, adding a one-degree tilt or Gaussian blur made the benchmark publisher’s provider perform much worse than competitors. He recommends thoughtful LLM-as-judge rubrics and testing models on customer-specific data rather than relying on generic benchmarks.
- Ng expects continued demand for AI infrastructure, citing potential applications across finance, healthcare, and transportation; he believes built inference and data-center capacity will be used, while acknowledging possible overbuilding and investor losses.
- He says software-engineering job postings are rising as AI makes engineers more productive, and that his teams will not hire candidates who cannot use coding agents; he identifies upskilling as a major challenge.
- Ng says chatbots can improve homework results while reducing learning and closed-book exam performance; he is leading Learn Vector, which he describes as a Coursera-funded team developing AI for learning and scalable learning experiences.
- For sovereign AI access, he favors open-weight models as an economical way to prevent outside parties from cutting off access. He also says AI has strengthened both cyber offense and defense, citing agent swarms and the open-source OpenWorker framework used by cyber defenders.
- Ng alleges that companies with large investments in proprietary models have lobbied to restrict freely accessible open models, and that AI-risk narratives can serve business, PR, and fundraising goals.
Parallel, which describes itself as building web-search APIs for AI agents, used its Task API to develop its first major out-of-home campaign: all six compositions drew on real research runs, with API responses incorporated into the artwork. The Task API returned web research as structured summaries and key facts with source links. The campaign is rolling out across New York, the San Francisco Bay Area, and London this fall and winter.
Parallel says its search and research APIs power millions of agent searches daily; its Search API retrieves fresh web evidence for cited answers. Parag Agrawal praised the team’s “deep obsession” and execution under extreme time constraints, saying this reflects how @p0 operates across disciplines.
- Fei-Fei Li said she co-founded World Labs in 2024 with three others; she described the company as having 40 people at the interview.
- World Labs is building a frontier model for “spatial intelligence,” which Li views as AI’s next frontier beyond language: capabilities for interacting with and designing in 3D environments, plus simulation and training. The company targets creators, robotics, healthcare, education, and enterprise users. Li grounds the thesis in her Stanford research in AI, computer vision, and robotics, and describes intelligence as encompassing perception, touch, planning, and reasoning in complex 3D worlds.
- AI inference could favor renewed hardware specialization: voltage scaling broke down around 2005, transistor costs stopped falling around 2012, and compute outpaced memory bandwidth. A 2026 Goldman Sachs report cited in the article projects token demand will rise 24× from 2026 to 2030, while available power increases only about 2×.
- The shift is not necessarily a near-term GPU replacement: the article notes scarce fab and HBM allocations, multi-year economics for deployed accelerators, and GPUs’ ability to absorb useful inference features. It expects more specialized inference chips alongside GPUs, while allowing that the longer-term pendulum may swing back toward general-purpose architectures.
- The proposed opportunity is heterogeneous inference orchestration—matching workload phases to GPUs and near-memory accelerators such as Groq, Cerebras, and d-Matrix. Gimlet’s article presents disaggregated approaches as delivering 2–10× faster performance within the same power envelope, but says deployment requires complex, modular hardware and software; physical-infrastructure complexity could be a durable moat.
Garry Tan highlighted Greptile, Blacksmith, and PostHog in response to a post captioned “Fastest growing startups per Brex data”; the supplied text gives no growth figures or specific ranking for the companies.
USOPM’s CIO office says it is transforming retirement services, consolidating 120+ HRIT systems, reimagining USAJobs, and targeting healthcare fraud, waste, and abuse. The office is hiring two GS-15 enterprise architects for senior, non-supervisory technical roles, framing the effort around a tech-first organization and improved developer experience; the roles are based in Washington, DC, Macon, GA, or Boyers, PA, and the posting opens October 22 and closes October 28.
Stepful (YC S21), which trains people for healthcare jobs to help address staffing gaps, trained 30,000 students last year and works with 39 healthcare systems, including seven of the 10 largest in the US; it recently raised a $55M Series C. The company has expanded from helping individuals become medical assistants to working with large health systems, and is using AI to make training more personalized and scalable.
Dalton Caldwell flags an email-attribution ambiguity: when a message seems strange, it may be unclear whether the person or their agent sent it.
YC-backed Blacksmith (W24), which provides cloud infrastructure for building and testing code, raised a $66M Series C at a $1B valuation; founder Adi Jayaprakash says Sequoia is leading the round. Blacksmith’s CI volume grew from 5 million jobs per week in January to more than 90 million, and over 10,000 companies now use the platform; YC describes code validation as an emerging bottleneck to shipping software.
The Hacker News flagged CVE-2026-110264 as a critical GitLab flaw (CVSS 10.0) affecting 12 years of releases and potentially enabling unauthenticated remote code execution on self-managed servers.
Legora (YC W24) raised a $550M Series E at an $8.8B valuation. Its agentic legal-work platform serves more than 140,000 legal professionals and their AI agents; ARR grew from $1M to $100M in under 18 months, then added another $100M in less than six months. The US is now Legora’s largest market by revenue, and more than half of the AmLaw 50 are customers.
- Guo says CVEs rose from ~4,000/month to ~15,000 since April. Although most newly disclosed CVEs are not yet exploited, she warns that applying a frontier-class model to the vulnerability backlog could close the “defenders window” and calls for a large reallocation of resources to cyber this year—not 10%.
- Guo characterized an NSA warning as telling PHP operators to patch two AI-found RCEs immediately or take servers offline. The embedded NSA notice confirms two critical PHP vulnerabilities and advises an update or taking systems offline if an immediate update is not feasible, but does not itself attribute their discovery to AI.
- @gdb, identified in the No Priors episode post as an OpenAI co-founder/president, said security spending should “vastly outstrip” software spending.
- Remediation capacity is under pressure: HackerOne customers halved fix times over two years, but their backlog grew 131% amid a 3x jump in intake. Oracle’s quarterly patch drop was 1,283, up from 374 a year earlier; Guo identifies hand-run regression suites with the same teams and tools as 2024 as a bottleneck.
- Guo’s proposed direction is agents reproducing, patching, testing, and shipping fixes, with engineers focused on the small number of cases requiring judgment. She raises formal verification, safe distributions, and how to protect organizations without red teams as unresolved questions.
Scott Kupor supports data-center providers partnering with local communities to address energy and water costs and noise restrictions, and to train residents for data-center jobs expected to continue after construction workers leave.
Martin Casado called the cyber situation “total fucking bedlam” and said security and operations teams needed support. In a post he shared, Eito Miyamura cited numerous security incidents and recommended updating software and active code repositories; it said uv 0.13.4 had been updated in response to an eMudra incident involving uv and urged refreshing package managers including uv, npm, bun, and cargo.
