AI Bug Flood: Microsoft Ships 1,461 Patches [2026]

Microsoft shipped 1,461 patches in October 2026, more than it released in all of 2025, and Google Chrome fixed 1,072 bugs since June, more than the prior two years combined, according to figures from a New York Times report published October 28, 2026 and relayed by Startup Fortune. The Times report, headlined "Silicon Valley Is Grappling With a Flood of Bugs Found by A.I.," describes AI agents reading large codebases at machine speed and flagging suspect code by the thousand, shifting the security bottleneck from finding flaws to verifying, reproducing and patching them. The same relayed figures include more than 1,300 open-source flaws found in OpenAI code since May, 792 critical and 10,500-plus high-severity findings plus 14 CVEs in Codex Security's first month, 661 patches written and 458 accepted for OpenAI Daybreak, a more than fourfold one-month rise in Node.js reports to the OpenJS Foundation, and 82% of more than 2,000 surveyed researchers using AI in bug-hunting workflows per Bugcrowd's 2026 report.

Silicon Valley has a new bottleneck, and it is not finding software flaws. AI tools now surface vulnerabilities faster than engineers can read, triage, and fix them. The New York Times reported on October 28, 2026 that AI has uncovered a horde of bugs, and that security teams are racing to patch them before attackers can abuse them. The AI-found bugs story has since spread through trade press, and the numbers behind it are hard to ignore. Microsoft shipped 1,461 patches in October alone, more than it released in all of 2025, according to figures relayed by Startup Fortune from the Times report. Google Chrome fixed 1,072 bugs since June, more than the prior two years combined. Those are not marginal increases. They describe a software industry whose defect backlog just got a spotlight the size of a data center. Our earlier look at AI-generated code security scores https://shattered.io/ai-generated-code-security-veracode-claude-opus-5-5-gpt-6-sol-73-percent-2026/ covered the other half of this equation, the code models write rather than the code they audit. This analysis walks through what was reported, which figures are firm, where the pressure lands first, and what developers and maintainers can do about it. It also marks clearly where the evidence stops and informed judgment begins. Where a number did not come from a named outlet, it does not appear here. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What the New York Times Reported About AI-Found Bugs The Times story, published October 28, 2026 under the headline “Silicon Valley Is Grappling With a Flood of Bugs Found by A.I.”, frames a simple shift. For decades, the scarce resource in software security was discovery. Skilled researchers found flaws slowly, and vendors fixed them at a pace that roughly matched. AI agents broke that balance by reading enormous codebases at machine speed and flagging suspect code by the thousand. Several outlets summarized the report in the days after. Startup Fortune, SBS News, and BigGo Finance all carried versions, and each described the same core pressure: engineers are in a sprint with no finish line. The detail that matters most is the direction of the constraint. Finding a bug now costs less than verifying it, reproducing it, writing a fix, testing the fix, and shipping the update to users. That inversion explains why the story landed as a headline rather than a footnote. A vendor that finds ten times more bugs does not automatically become ten times safer. It becomes ten times busier, and the gap between disclosure and deployment of a patch is exactly the window attackers want. The Numbers Behind the Flood Specific figures give the story its weight. The table below collects the numbers reported by named outlets and attributes each to its source. Treat the Times-derived items as relayed figures, since they reached this article through secondary coverage rather than the original text. | Metric | Figure | Reported by | |---|---|---| | Microsoft patches, October 2026 | 1,461 more than all of 2025 | New York Times, via Startup Fortune | | Chrome bugs fixed since June | 1,072 more than the prior two years combined | New York Times, via Startup Fortune | | OpenAI open-source flaws found since May | More than 1,300 | New York Times, via Startup Fortune | | Codex Security first-month findings | 792 critical, 10,500+ high-severity, 14 CVEs | Startup Fortune | | OpenAI Daybreak patches written / accepted | 661 / 458 | CyberScoop | | Node.js OpenJS Foundation report volume | More than fourfold in one month | Startup Fortune | | Researchers using AI in hunting workflows | 82% of 2,000+ surveyed | Bugcrowd 2026 report, via Startup Fortune | Read the table as a set of different measurements rather than one trend line. Microsoft’s count measures shipped fixes, which reflects both discovery and the capacity to respond. OpenAI’s figures measure what its own tools surfaced. The Bugcrowd survey measures researcher behavior. They point the same way, but they do not add up to a single industry-wide total, and no outlet has published one. Microsoft and Chrome Set the Pace The two largest numbers in the Times coverage come from companies with mature security engineering. Microsoft already runs one of the biggest patch pipelines on earth. Pushing 1,461 fixes in a single month, more than the full prior year, means the pipeline itself had to scale, not just the discovery side. Chrome’s 1,072 fixes since June tell a similar story over a longer window. Browsers sit at the front line because they parse untrusted content all day, and every flaw in one can reach a very large user base. A count that beats the previous two years combined suggests the discovery rate stepped up in 2026 rather than drifting upward. One caution applies. A raw patch count does not say how severe the fixes were. Some of the 1,461 are likely low-impact, and the coverage does not break them down by severity. Anyone claiming Microsoft became that many times less secure or more secure from the count alone is reading beyond the data. What the count does show is operational load, and that load now competes with feature work for the same engineers. How OpenAI’s Codex Security Feeds the Pipeline OpenAI sits on both sides of this story. According to Startup Fortune, its Codex Security product launched in March 2026 after an earlier research preview called Aardvark. In its first month, the tool flagged 792 critical issues and more than 10,500 high-severity issues, and 14 of the findings earned CVE identifiers. The Times report adds that OpenAI has found more than 1,300 flaws in open-source software since May. Daybreak and the cost of coordination CyberScoop covered remarks by Michael Aiello, OpenAI’s cyber executive in residence, at the DataTribe Challenge in Washington, D.C. on October 27. He said OpenAI’s Daybreak effort developed and tested 661 patches, and maintainers accepted 458 of them. That works out to about 69 percent accepted, a figure derived from the two reported numbers. Aiello also described the shape of the problem for maintainers. A volunteer who receives hundreds of AI-submitted vulnerability reports will often stop responding, and some projects stop accepting reports at all. OpenAI’s advice, as CyberScoop summarized it, is to file only high-quality reports, make each one readable by a single person, spend your own resources coordinating disclosure, and work across the whole ecosystem. The PHP Foundation case, where OpenAI helped validate findings, build fixes, and coordinate a release, was his example, though he acknowledged the execution was imperfect. The takeaway is that discovery vendors are starting to accept responsibility for the downstream cost. That is a meaningful change in norms. The cheapest report to send is the most expensive one to receive. Open Source Maintainers Take the Hardest Hit Large vendors can hire. Volunteer maintainers cannot. The OpenJS Foundation, which stewards Node.js among other projects, saw bug reports rise more than fourfold in a single month once AI scanning tools arrived, per Startup Fortune. A project with a handful of part-time reviewers does not absorb a fourfold jump without dropping something. Linux kernel and the duplicate-report problem On May 17, 2026, Linus Torvalds addressed the Linux kernel mailing list about AI-assisted reports, as Help Net Security detailed the next day. The complaint centered on duplication. When many researchers run the same tools against the same code, the same flaw reaches the same inbox over and over. Help Net Security also noted reports that lack a proof of concept and theoretical scenarios that fall apart under review. Delana Technologies, citing Security Boulevard, reported that kernel CVEs per release climbed from roughly 500 in the 6.x era to more than 1,000 at Linux 7.0, more than 1,500 at Linux 7.2, and nearly 2,000 approaching Linux 7.3. That second-hand figure deserves some caution, but the direction matches everything else in the record. curl’s experiment with no money The curl project offers the clearest natural experiment. Per Help Net Security, curl eliminated monetary rewards on January 26, 2026 and returned to HackerOne on February 25. Lead developer Daniel Stenberg observed on April 22 that report quality improved after the payouts ended. Fewer low-quality submissions arrived, and the share of confirmed vulnerabilities went up. Money, in other words, was drawing the noise. Bug Bounty Programs Start Closing Their Doors Bounty programs were built for a world where a human researcher put hours into each report. Automated submission breaks that economic model. A tool can file a hundred plausible-looking reports for the price of one, and the program pays the review cost. | Program | Change | Date | Source | |---|---|---|---| | curl | Dropped monetary rewards, returned to HackerOne | Jan. 26 and Feb. 25, 2026 | Help Net Security | | GitHub bug bounty | Requires validated submissions with a working proof of concept | Reported May 2026 | Help Net Security | | Intel | Replaced paid bounty $500 to $100,000 per flaw with unpaid disclosure on Intigriti | Around Sept. 19, 2026 | Tom’s Hardware, via Delana Technologies | | Google OSS VRP | Paused product vulnerability submissions | Effective Oct. 1, announced Oct. 5, 2026 | Help Net Security | Google’s move drew the most attention. The Open Source Software Vulnerability Reward Program launched in 2022 and covers projects including Go, Angular, and Protocol Buffers. Google said it saw a significant rise in automated submissions, and that the vast majority of them are not valid. Reports filed before October 1 are unaffected, some Cloud VRP coverage continues, and researchers can turn to other reward programs or the Patch Rewards Program. Google promised an update in the first quarter of 2027. Intel gave no public reason for its change, and Delana Technologies stressed that any link to AI-assisted reporting is observer speculation. Still, the timing across four programs in nine months makes the pattern hard to dismiss. Why Real Bugs and AI Slop Arrive Together The flood is not one phenomenon. It is two, tangled together. Part of it consists of genuine vulnerabilities that capable models find in real code. Part of it consists of hallucinated or unexploitable reports that read well and fall apart on inspection. Maintainers must open both envelopes to tell which is which. That is why OpenAI’s 69 percent patch acceptance rate and Google’s complaint about invalid reports can both be true. A well-run internal pipeline with human validation produces patches that maintainers want. An open submission queue attracts anyone with a model and a reward incentive. The tool is the same. The process around it decides whether the output helps. Duplication adds a third layer. Torvalds’s complaint was about many people reporting the same real issue. Nothing in that case is false, yet the effort wasted is nearly identical to handling a fake report. Coordination, not accuracy, is the missing piece there. Our coverage of GLM 5.3 cracking six Firefox builds at $1,000 each https://shattered.io/epoch-ai-glm-5-3-firefox-patch-gap-six-versions-1000-each-2026/ shows how cheaply a determined party can run these tools at scale. The Race Between Finding and Exploiting Defenders are not the only ones with AI. Delana Technologies described a case that shows how short the window has become. A flaw in Rejetto HFS, tracked as CVE-2026-61500 with a CVSS score of 9.8, was published on September 30, 2026. Horizon3.ai and The Register reported that exploitation was detected on the evening of October 1, about a day later. Delana attributes the original discovery to an AI system called Mythos. The same account puts roughly 11 weeks between the CVE record on July 13 and the first exploitation. So the system did not move instantly. It moved slowly through disclosure and then quickly once details went public. For defenders, the lesson is that the interval between publication and attack can collapse to a day, and any patching process that takes weeks leaves a large target. Attackers face the same triage problem defenders do, but they only need one working exploit. Nothing in the reporting says AI has made exploitation easier in general. It does say that the price of looking for exploitable flaws has dropped, which raises the odds that someone looks. For a view of how far offense testing has come, see our piece on Anthropic models hacking three organizations in capture-the-flag tests https://shattered.io/anthropic-ai-models-hacked-3-organizations-capture-the-flag-tests-2026/ . Fixing Is Harder Than Finding A finding is a claim. A fix is a change to production code that must not break anything else. That asymmetry is the heart of the backlog. StackHawk CEO Joni Klippert told the Times, as relayed by Startup Fortune, about a financial services company whose monthly code output jumped from 25,000 lines to 250,000 lines with AI help. About one million lines went unreviewed as a result. That anecdote points at a second-order effect. AI that writes code more quickly produces more code to audit, and AI that audits code more quickly produces more findings to fix. The two curves compound. Teams that adopt coding assistants without scaling review capacity create the very backlog the Times describes. Startup Fortune also pointed to a February report from The Register on the quality of AI-generated fixes. The details were not summarized in the coverage this article drew on, so the safe reading is narrow: automated patches still need human review, and some arrive wrong. That matches the experience of any engineer who has merged a machine-written patch that passed tests and still missed the root cause. Our walkthrough of five vibe coding safety practices https://shattered.io/vibe-coding-safety-practices-engadget-five-2026/ covers the review habits that matter most. Market Impact: Security Spend, Tooling, and Engineering Time No outlet in the sources reviewed here has published a dollar figure for the industry cost of the flood, so this section stays qualitative. Three effects are visible from the reporting. - Triage tooling gains value. When discovery is cheap, vendors that deduplicate, rank, and validate findings become the bottleneck breakers. Expect more products that pair a finding with a working proof of concept, since GitHub now demands exactly that. - Bounty economics get rewritten. Four programs changed terms in nine months. Pay-per-report models face pressure to move toward verified-exploit pricing, invitation-only access, or no cash at all, as curl chose. - Patch capacity becomes a competitive trait. A vendor that ships fixes quickly and safely gains trust. One that drowns in its backlog risks the 11-week gap that Rejetto HFS showed. Engineering time is the quiet cost. Every hour a Chrome or Windows engineer spends on a fix is an hour not spent on features. Public companies rarely disclose that tradeoff, yet the patch counts above imply it. Teams reading the same signals in cloud infrastructure should compare notes with our report on the Azure Kubernetes configuration flaw rated CVSS 10.0 https://shattered.io/azure-kubernetes-configuration-service-cve-2026-83938-cvss-10-2026/ , where speed of response mattered as much as the flaw itself. Competitive Comparison: How Major Players Respond Each organization in the story chose a different defense. The table sets those choices side by side, using only actions that the cited outlets reported. | Organization | Stance | Reported action | |---|---|---| | Microsoft | Absorb and ship | 1,461 patches in October 2026 | | Google Chrome | Absorb and ship | 1,072 bug fixes since June | | Google OSS VRP | Pause intake | Product vulnerability submissions halted Oct. 1, update due Q1 2027 | | OpenAI | Find, validate, coordinate | Codex Security, Daybreak patches, 458 of 661 accepted | | Intel | Remove cash incentive | Paid bounty replaced by unpaid Intigriti program | | curl | Remove cash incentive | Rewards dropped Jan. 26, quality rose by Apr. 22 | | GitHub | Raise the bar | Working proof of concept required | The split is revealing. Companies with large engineering staffs absorb the volume. Programs that depend on outside reporters raise the cost of entry, either by cutting pay or by demanding proof. OpenAI stands apart because it sells the discovery tool and also funds the downstream work to make its findings usable. Historical Context: From Fuzzers to Language Models Automated bug finding is not new. Fuzzers have hammered open source code for years, and they also produced piles of crashes that maintainers had to sort. Every earlier wave eventually settled once projects built filters, templates, and rules for what counts as a valid report. The 2026 wave differs in two ways. First, language models write plausible reports, complete with narrative and suggested severity, which makes the junk harder to dismiss at a glance. Second, the price of running a scan fell low enough that many parties now scan the same code at once. A fuzzer needed setup and compute. A model prompt needs a subscription. That history suggests a settling period is coming. It also suggests the settling will cost some projects their volunteers first. Supply-chain visibility helps in the meantime, because you cannot patch what you cannot inventory. Our SBOM tutorial using Syft, Grype, and Cosign https://shattered.io/sbom-tutorial-syft-grype-cosign-12-steps-2026/ walks through building that inventory in 12 steps. What Developers and Security Teams Should Do Now The practical response splits by role. Maintainers need ways to cut intake load. Product teams need patch pipelines that scale. Researchers need to stop adding to the noise. - Maintainers: require a reproducible proof of concept, publish a short template for reports, and consider ending cash rewards if noise outweighs signal, as curl did. - Product teams: rank findings by exploitability and exposure, not by raw count. Track time from disclosure to deployed fix as a core metric. - Researchers using AI: validate each finding by hand before filing, check for existing reports, and follow OpenAI’s stated advice to coordinate disclosure across the ecosystem. - Everyone: keep an up-to-date software inventory so a new CVE maps to your systems within minutes. Teams should also watch for flaws already marked as exploited. The CISA catalog remains the best signal for what to patch first, and our piece on the VMware ESX flaw rated just 2.7 yet listed on CISA KEV https://shattered.io/vmware-esx-cve-2026-41709-cvss-2-7-cisa-kev-warp-panda-2026/ shows why a low score can mislead. Reviewing the full Software section https://shattered.io/software/ for developer tooling updates helps teams keep pace. Five Predictions for the Next 12 Months The following predictions are this site’s analysis, not reported facts. They extend the trends above and may prove wrong. 1. More bounty programs will narrow or close. Four changes in nine months suggests the trend has momentum. Expect at least a few additional large programs to move to invitation-only or proof-required models. 2. Verified-exploit submissions will become the norm. GitHub already demands a working proof of concept. Others will follow, because validation is the only filter that scales. 3. Triage and dedupe tools will draw investment. The bottleneck moved from discovery to review, and capital follows bottlenecks. 4. Disclosure norms will formalize. OpenAI’s coordination advice hints at an emerging standard where discovery vendors share the downstream burden. Industry coalitions will likely codify it. 5. Patch counts will keep setting records. Microsoft’s and Chrome’s totals are unlikely to fall soon, which means “more patches” will stop being a story and “time to deploy” will become the metric that matters. The Takeaway for Engineers The AI-found bugs flood is real, measurable, and still poorly understood in aggregate. The Times reporting and the follow-up coverage give solid anchors: 1,461 Microsoft patches, 1,072 Chrome fixes, more than 1,300 OpenAI-found open source flaws, and a string of bounty programs that changed terms. What they do not give is an industry-wide cost or a clear split between genuine bugs and noise. The takeaway: finding vulnerabilities got cheap, and fixing them did not. Organizations that treat the flood as a discovery problem will buy more scanners and fall further behind. Those that treat it as a throughput problem will invest in validation, deduplication, and deployment speed. The Rejetto HFS timeline, a day from publication to exploitation, shows what happens to the teams that wait. FAQ: AI-Found Bugs and the Patch Backlog What did the New York Times report about AI-found bugs? On October 28, 2026, the Times reported that AI has uncovered a large number of software flaws and that Silicon Valley engineers are racing to fix them before hackers exploit them. Coverage cited 1,461 Microsoft patches in October and 1,072 Chrome fixes since June. Why are open source maintainers struggling? Maintainers receive far more reports than they can review. Help Net Security described heavy duplication, missing proofs of concept, and theoretical scenarios that fail on inspection. The OpenJS Foundation saw Node.js reports rise more than fourfold in a month, according to Startup Fortune. Did Google really pause its open source bug bounty? Yes. Help Net Security reported that Google paused product vulnerability submissions to its OSS VRP effective October 1, 2026, citing a rise in automated submissions that are mostly invalid. Earlier reports are unaffected, and Google promised an update in the first quarter of 2027. Are all AI-found bugs fake? No. OpenAI’s Daybreak effort wrote 661 patches and maintainers accepted 458, per CyberScoop. Genuine findings exist alongside low-quality reports, and the challenge is sorting them quickly. Why did curl drop its cash bounty? curl eliminated monetary rewards on January 26, 2026, and returned to HackerOne on February 25. Lead developer Daniel Stenberg noted on April 22 that report quality improved, per Help Net Security. How fast can attackers exploit a newly published flaw? In one reported case, exploitation of Rejetto HFS CVE-2026-61500 was detected about a day after the September 30 publication, according to Horizon3.ai and The Register as summarized by Delana Technologies. What should a development team do first? Build a software inventory, rank findings by exploitability, and track time from disclosure to deployed fix. Require human validation of any AI-generated finding or patch before it ships.