Verified timeline
| Date | Event | Source | Certainty |
|---|---|---|---|
| August 11, 2026 | Security firm Calif reported the first PHP vulnerability discovered using OpenAI's GPT-5.6 Sol model | Al Jazeera | Reported by outlets |
| August 24, 2026 | Wiz security firm reported the second vulnerability | Al Jazeera | Reported by outlets |
| Tuesday (October 21) | Vulnerability details leaked online, a day before planned patch release | France 24 | Reported by outlets |
| Wednesday, approximately 1345 GMT | PHP developers released a new version fixing the vulnerabilities | CNBC | Reported by outlets |
| Wednesday (October 22) | Government websites across Europe went offline as precautionary measure; NSA advised organizations to patch immediately or take systems offline | Al Jazeera | Stated by officials or the organisation concerned |
| Late Wednesday | Cloudflare CEO reported millions of exploitation attempts after public disclosure, jumping from thousands before announcement to millions within hours | France 24 | Reported by outlets |
The sources differ on the estimated percentage of websites using PHP: CNBC cites roughly 75 percent, while France 24 and Al Jazeera cite about 70 percent.
Why this matters
PHP powers approximately 70-75% of websites whose programming language can be identified, according to CNBC. The two critical vulnerabilities discovered through AI allow attackers to "remotely run commands on a server running old PHP code" without authentication, according to Al Jazeera. France 24 reported that Cloudflare saw millions of exploitation attempts after public disclosure, with thousands of attempts occurring before the announcement and jumping to hundreds of thousands within two hours.
Government websites across Europe and Asia went offline as precaution. CNBC reported sites in the Netherlands and Finland went dark, while France 24 documented shutdowns in France, Belgium, and the Czech Republic. The U.S. National Security Agency advised organizations to patch immediately or take vulnerable systems offline, according to Al Jazeera.
Small websites on shared hosting face particular risk. CNBC noted such sites lack technical resources to patch quickly and may remain exposed longer after fixes become available. The PHP Foundation warned that AI tools now find security bugs faster than humans can patch them.
What is established, alleged and stated
Reported by the outlets
- According to CNBC, PHP developers identified two critical vulnerabilities and released a fix on Wednesday at approximately 1345 GMT, with reports that hackers were exploiting the flaws before the patch became available.
- France 24 reports that vulnerability details leaked online Tuesday, a day before the planned fix release, prompting government agencies to take websites offline as a precautionary measure.
- According to Al Jazeera, the two vulnerabilities allow attackers to remotely run commands on vulnerable servers without password or account access.
- CNBC reports that Cloudflare observed "early attacks emerging" and was updating firewall rules, while France 24 states that Cloudflare saw millions of exploitation attempts within hours of public disclosure.
Stated by officials or the organisation concerned
- According to France 24, the PHP Foundation stated the vulnerabilities were "among the most severe vulnerabilities discovered in PHP, and arguably in software history."
- CNBC reports Automattic stated it was "working with the PHP security team to ensure a smooth update."
Alleged or disputed
Nothing in the sources.
What we still do not know
-
CNBC reports that officials said "all PHP versions from 5 onward could be affected," but PHP had not yet disclosed which versions are actually vulnerable.
-
The sources differ on when vulnerability details became public. France 24 states details leaked Tuesday, one day before the planned patch release, but CNBC does not mention this leak.
-
The reports do not specify how many websites were successfully compromised versus merely targeted by exploitation attempts.
-
CNBC notes the vulnerability "could potentially open the door to millions of servers" but does not estimate actual breach numbers.
-
The sources do not establish when government and corporate websites taken offline as precautions returned to service.
-
Al Jazeera states the flaws "had been present for decades" but does not explain why they remained undiscovered until AI tools identified them.
What the models agree on
- AI tools discovered two critical vulnerabilities in PHP, a widely used programming language.
- The vulnerabilities allow attackers to execute remote commands without authentication on vulnerable servers.
Agreement between models does not validate an allegation. It only shows that they read the same reports in the same way.
Where they differ
Every difference is typed. An omission is not treated as an error: it is only called critical when leaving the point out would change how a reader understands the story.
| Divergence | Type | Why it matters | Models concerned |
|---|---|---|---|
| A PHP security patch was released on Wednesday at approximately 1345 GMT. | Critical omission | Establishes when the fix became available to organizations. | Left out by Mistral and Grok |
| The NSA advised organizations to patch PHP immediately or take vulnerable systems offline. | Critical omission | Establishes official U.S. government response and guidance to the vulnerability threat. | Left out by Claude |
| Hackers were actively exploiting the vulnerabilities before the security fix was released. | Critical omission | Indicates attackers had access to exploit the vulnerabilities during the window before patching. | Left out by GPT, Gemini, Mistral, Perplexity and Grok |
| Vulnerability details leaked a day before the planned patch release, prompting pre-emptive shutdowns. | Critical omission | Explains why governments took websites offline before the official patch release date. | Left out by Claude, GPT, Gemini, Mistral and Perplexity |
| Government websites in multiple European countries went offline as a precautionary measure. | Different emphasis | Demonstrates the widespread impact of the vulnerabilities on critical infrastructure. | GPT (general terms only) |
| Cloudflare observed millions of exploitation attempts following public disclosure of the vulnerabilities. | Different emphasis | Demonstrates active exploitation of the vulnerabilities after details became public. | GPT (general terms only) |
| PHP powers approximately 70-75% of websites whose programming language can be identified. | Neutral omission | Provides context on PHP's prevalence and therefore the scale of potential impact. | Left out by Gemini and Perplexity |
| PHP Foundation or security firms called the vulnerabilities among the most severe in PHP's history. | Neutral omission | Emphasizes the severity assessment from authoritative sources in the PHP ecosystem. | Left out by Claude, GPT, Gemini and Perplexity |
| Security firm Calif discovered the vulnerabilities using OpenAI's GPT-5.6 Sol model in August. | Neutral omission | Identifies the initial discoverer and specific AI model used in the discovery process. | Left out by GPT, Mistral, Perplexity and Grok |
| Wiz security firm independently discovered vulnerabilities weeks after Calif using AI models including Google's Gemini. | Neutral omission | Documents that multiple security firms independently discovered the same vulnerabilities. | Left out by GPT, Mistral, Perplexity and Grok |
Model reliability scorecard
| Model | Coverage | Critical omissions | Unverified claims | Contradictions | Editorial reliability |
|---|---|---|---|---|---|
| Claude | 75% | 2 | 0 | 0 | Solid |
| GPT | 58% | 2 | 1 | 0 | Weak |
| Gemini | 67% | 2 | 0 | 0 | Solid |
| Mistral | 58% | 3 | 0 | 0 | Mixed |
| Perplexity | 50% | 2 | 0 | 0 | Mixed |
| Grok | 67% | 2 | 1 | 0 | Mixed |
How the rating works. Coverage is the share of the panel's points that a model includes. Each model was asked for 6 to 10 claims, so coverage reflects what it chose to include within that budget. Penalty points: 1 per critical omission, 2 per unverified claim, 2 per contradiction, 1 if coverage is below 60%. Strong: coverage of 80% or more and no penalty. Solid: coverage of 65% or more and at most 2 penalty points. Mixed: at most 4 penalty points. Weak: more than 4. There is no numerical grade.
The most important finding
The main result is not that the models disagree. None of them contradicted another. They differ mostly by omission: 4 critical and 4 neutral across 12 points, plus 2 differences of emphasis. The more important signal is reliability: GPT and Grok added 2 claims that the cited reports do not support, and those matter more than any omission.
Reading between the lines
What this suggests: Al Jazeera reports that the underlying flaws "had been present for decades" yet required "a chain of steps at the edge of what human researchers could practically achieve" to exploit. This implies AI's contribution was not discovering a new weakness but demonstrating exploitability at scale. If this reading holds, it would mean the vulnerabilities existed within reach of sophisticated human attackers for years before AI exposure—a fact that would support urgency around patch adoption going forward.
What this does not establish: CNBC and France 24 confirm that Cloudflare saw "millions" of exploitation attempts after disclosure, but neither source specifies how many attacks succeeded in compromising systems or extracting data. Active attempts and successful breaches are distinct events the reporting does not resolve.
The key signal: France 24 notes that vulnerability details "leaked online Tuesday, a day before the planned release of the fix," forcing premature shutdowns. This timing gap suggests that rapid disclosure windows—whether accidental or engineered—may now constrain defenders' ability to patch methodically, amplifying pressure on organizations with limited technical capacity.
What to watch
-
Will small websites and shared hosting providers patch quickly enough to prevent widespread exploitation? CNBC reported that small sites on shared hosting services could "take longer to patch and remain exposed well after fixes become available," creating a window for attackers.
-
How many exploitation attempts will occur now that vulnerability details are public? France 24 reported that Cloudflare CEO Matthew Prince observed "millions" of exploitation attempts within hours of public disclosure, after "thousands" in the 10 hours before announcement.
-
Which PHP versions will PHP officially confirm as vulnerable? CNBC noted that officials said "all PHP versions from 5 onward could be affected," but PHP had not yet disclosed which versions are actually vulnerable.
-
Will the PHP Foundation's warning about AI-accelerated vulnerability discovery prompt coordinated industry responses? Al Jazeera reported the foundation called for "a coordinated response to avoid major disruptions" as AI tools find security weaknesses faster than human programmers.
-
How quickly will government websites taken offline return to service? France 24 reported websites in multiple countries went dark as a precautionary measure, leaving unclear timelines for restoration after patching.
Planned review of this page: October 29, 2026.
Method, honestly stated
- Every model received the same reports (CNBC, France 24, Al Jazeera) and the same prompt, and was told to use only those reports.
- Models that ran: Claude, GPT, Gemini, Mistral, Perplexity and Grok. No model was left out.
- Each claim a model made was checked against the reports. A claim counts as verified if its supporting passage appears word for word in a report, or if a separate model judged it a supported paraphrase. Otherwise it is marked unverified.
- A separate model merged equivalent claims into comparable points and classified each difference as contradiction, critical omission, neutral omission, different emphasis or unverified addition. That step is automated and can be wrong.
- The similarity score (91/100) measures how alike the analyses read. It is not an accuracy or confidence score, not agreement on facts and not a measure of truth.
- An automated pipeline wrote and published this page. No person read it before publication.
Verdict
Coverage of the central fact is broadly consistent: 2 of 12 points appear in every analysis. The differences are mostly about completeness and framing, not contradiction. The main risk identified is added claims: GPT and Grok included material the cited reports do not support.
