Claude Flags 493 Issues as OpenAI and Google Claim PHP Find

Anthropic says its Claude model found 493 issues across 25 open-source projects covering more than five million lines of code, in a research post dated 20 October 2026. One conflicting pair of settings it flagged “had been present in the software for 21 years.” A day later, on 21 October 2026, the PHP project shipped fixes for two critical flaws that OpenAI and Google both say their models helped find.

OpenAI says security firm Calif “identified a critical flaw using GPT‑5.6 Sol” under its Patch the Planet initiative. Security firm Wiz says it found both flaws using its own tool, Wiz Atlas, together with Google’s Gemini 4 Argon and GPT-5.6 Sol, and that Gemini 4 Argon “made an invaluable contribution to finding the vulnerability.” Each claim below is the company’s own.

Business Pill · THE AI AUDIT PROBLEM

A short Business Engineer explainer on why checking AI-made work can cost more than making it. It teaches the general idea only and says nothing about any company in this piece.

What Anthropic Says Claude Found

AI vulnerability hunting means pointing a model at a codebase and asking it to find weaknesses an attacker could use. Anthropic’s post, by Chris Rohlf, takes a slightly different angle: instead of chasing single bugs, Claude looked for code that outsiders can reach but should not need to, which Anthropic calls attack surface. Anthropic argues that “reducing code reachability can reduce the severity of unknown vulnerabilities, including the ones that might be under active exploitation.”

Anthropic says Claude built “an inventory of over 17,000 configuration options” across the 25 projects, of which 3,842 related to a security-relevant control or feature, and found 493 issues. Claude also found 162 security-sensitive combinations of settings that were accepted without an error or warning, and Anthropic says that “In 94 of those cases, the combination silently turned off a security control.” For the gated projects, Anthropic says: “None of these issues would have been immediately exploitable on their own, but they create a much larger attack surface than is necessary.”

Its case study is OpenSSH, the widely used remote-access software. Anthropic says Claude found roughly 4,500 lines of new, experimental code that an unauthenticated client could reach in the default configuration, “about 16% of OpenSSH’s own pre-authentication code.” Anthropic says it reported this to the maintainers with patches, that the resulting fix “removed roughly a fifth of the OpenSSH code that an unauthenticated client previously could have reached in a default build,” and that the changes shipped in version 10.6.

What Was Posted on X

Chris Rohlf, who wrote Anthropic’s post, says the team tasked Claude with autonomously mapping and reducing the attack surface of 25 large open-source codebases, and that the work cut OpenSSH’s pre-authentication attack surface, now fixed in 10.6. His post says over 500 findings; the Anthropic post itself gives 493 issues, the figure this piece uses.

OpenAI’s post says its models helped researchers in its Patch the Planet initiative discover critical flaws in PHP, that it led a coalition of maintainers, researchers and security providers for two months before disclosure, and that every site using PHP can now upgrade. The claims are OpenAI’s own.

What OpenAI Says About PHP

OpenAI says it “led a coalition of maintainers, researchers, and security providers” over the past two months, and that Calif found a critical flaw using GPT-5.6 Sol as part of Patch the Planet, “with Wiz also independently discovering and reporting both vulnerabilities.” It credits PHP’s maintainers with leading the fixes and the release.

OpenAI calls the PHP case “one of the most impactful security issues in recent memory” and says that, to its knowledge, “frontier AI found these vulnerabilities before attackers did.” It says it has “committed $1 billion in subsidized Daybreak access” for defenders with limited security resources, and that these bugs are “among several significant vulnerabilities we’ve already found and are working to fix.” It also says it will launch “an accompanying Codex skill” to help operators check which PHP versions they run.

What Wiz and Google Say

Wiz says it used Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol throughout the research, and that Gemini 4 Argon also helped with exploit development, attack detection and validating PHP’s patch. Raluca Ada Popa, head of Gemini security at Google DeepMind, is quoted in Wiz’s post: “We’re pleased to see Gemini 4 Argon help defenders bridge that gap, finding vulnerabilities and helping patch them.” Google Cloud’s security bulletin GCP-2026-070 says: “While Google Cloud itself is not vulnerable, users of this language running on Google Cloud may be vulnerable.” It tells customers to “take immediate action to protect their workloads by redeploying their workloads with updated dependencies.”

PHP says its releases 8.5.12, 8.4.27, 8.3.36 and 8.2.35 “fix two critical vulnerabilities that can allow unauthenticated remote code execution,” tracked as CVE-2026-109618 and CVE-2026-109619. W3Techs says PHP is used by 69.3% of the websites whose server-side programming language it knows. Wiz says that with these tools “a single researcher completed in a week what previously would have taken a world-class team months of work.”

Three Labs, One Week, Side by Side

The table sets out what each company says, row by row. The rows are not comparable units: Anthropic counts issues found by shrinking attack surface, while the PHP rows are two critical flaws with CVE numbers. The table is our tally of the companies’ own statements.

Connecting the Coverage

Our PHP follow-up reported Wiz’s account that Gemini 4 Argon and GPT-5.6 Sol helped find the two flaws. Our first PHP piece quoted Cloudflare’s chief technology officer, Dane Knecht, saying before the release: “We are seeing early attacks emerging and are updating mitigations as new vectors show up.” OpenAI’s view, quoted above, is that AI found the flaws before attackers did; the two statements are each the speaker’s own.

Our AI-attacks angle piece covered the offence side, CrowdStrike’s report that a hacker targeting banks used Claude, Grok and DeepSeek. Our OpenShell piece covered Nvidia’s design for running AI agents under enforced limits, and our XBOW piece covered a security firm that won a Chrome bug bounty with GPT-5.6 Sol. As we read it, this week adds the defenders’ side at scale: three labs now publish what their models found.

The Structural Read

As we read it, bug finding is turning into a product line: OpenAI says it has committed $1 billion in subsidized Daybreak access for defenders, and Anthropic says it combined its techniques into one reusable Claude skill that it is working to integrate into OSS-Scanner.

As we read it, credit is now contested: OpenAI says Calif found a critical PHP flaw with GPT-5.6 Sol and that Wiz found both independently, while Wiz says Gemini 4 Argon made an invaluable contribution.

As we read it, the bottleneck moves to the people who confirm and ship fixes: OpenAI describes a staged response with maintainers, providers and government agencies, and Anthropic says it reported its OpenSSH finding to the maintainers with patches.

Three Implications

FOR SOFTWARE MAINTAINERS Anthropic says Claude found that all eleven gated projects in its study had code that was unintentionally reachable, and OpenAI says similar findings are expected to “increase sharply in the coming months.”

FOR COMPANIES RUNNING PHP PHP lists 8.5.12, 8.4.27, 8.3.36 and 8.2.35 as the fixed versions, and OpenAI says it is sharing a PHP version checker.

FOR AI LABS Anthropic, OpenAI and Google each now tie a named model to a security result, and OpenAI and Wiz describe the same PHP case with different emphasis.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework argues that the SKILL.md file “is becoming the modular unit of delegated work — the line of code of the agentic era.”

As we read it, both labs packaged their security work that way this week: Anthropic says it combined its techniques “into one reusable Claude skill,” and OpenAI says it will launch “an accompanying Codex skill” for PHP operators. The framework was published in July 2026 and does not discuss these findings.

What Is Not Established

Every finding here is described by the company that made it. Anthropic calls its 493 findings issues and attack surface, and says the gated-project issues were not immediately exploitable on their own; the 493 are not presented as confirmed vulnerabilities with CVE numbers. Credit for the PHP flaws is described differently: OpenAI says Calif found a critical flaw with GPT-5.6 Sol and Wiz found both independently, while Wiz highlights Gemini 4 Argon.

We do not adjudicate between those accounts. We did not contact Anthropic, OpenAI, Google, Wiz, Calif or PHP, and we deliberately leave out technical details of any flaw.

The Bottom Line

Within two days, Anthropic, OpenAI and Google each tied a model to a security finding: 493 issues for Claude, and two critical PHP flaws that OpenAI and Google both claim a share of. As we read it, the business point is that bug finding is becoming a product line for AI labs, sold through subsidies, skills and partners, and the bottleneck moves to the people who confirm and ship the fixes.