SafetyRisk, alignment & guardrails
AI-Discovered PHP Flaws Exploited Before Patch as Q3 Vulnerability Disclosures Triple
Researchers using AI found critical PHP flaws, and attackers began exploiting them before the fix shipped. Meanwhile, disclosed vulnerabilities roughly tripled year over year, and cyber agencies in the UK and India warned that attackers are adopting AI tools against infrastructure and finance.
From the WSJ · Technology article of November 6, 2026 and cited sources: In August, a Wiz team's AI agent found a long-standing PHP memory-handling bug while it was trying to exploit a separate Drupal weakness. [3] [4]
SOURCES[3] TIME: The AI-Discovered Bug That Could Have Become a… time.com[4] National CIO Review: AI-Discovered PHP Vulnerabilities nationalcioreview.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- In August, a Wiz team's AI agent found a long-standing PHP memory-handling bug while it was trying to exploit a separate Drupal weakness. [3] [4]
- Original source
- WSJ · Technology article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 6, 2026 article published
- Covered
- Nov 6 edition
- Evidence
- 11 sources across 10 sites
- 1The Morning Risk Report: Companies Face Cyber Pain as…wsj.com
- 2Insurance Journal/Reuters: AI Sharpens Cat-and-Mouse…insurancejournal.com
- 3TIME: The AI-Discovered Bug That Could Have Become a…time.com
- 4National CIO Review: AI-Discovered PHP Vulnerabilitiesnationalcioreview.com
- 5BleepingComputer: Binance links $6.3B theft to compromised…bleepingcomputer.com
- 6Insurance Journal: UK Must Expect AI-Powered Attack on…insurancejournal.com
Reach: not measured yet
In August, a Wiz team's AI agent found a long-standing PHP memory-handling bug while it was trying to exploit a separate Drupal weakness. Calif, working with OpenAI, found a related flaw on its own. Both flaws could allow unauthenticated code execution on servers. Researchers said this kind of work has shrunk from months to hours or days. [3] [4]
Calif posted an exploit video by mistake before the fix was ready, which pushed PHP to ship emergency patches on October 21. Exploitation had already started. Cloudflare's own telemetry showed attempts climbing from thousands to millions. Censys counted nearly 55 million web properties running PHP, though not all of them are confirmed vulnerable. [3] [4]
About 37,000 vulnerabilities were disclosed in Q3 2026, roughly triple the figure a year earlier. Claroty said it recreated an attack for about $6 in 28 minutes. Companies report serious flaws now arrive every few days, and one October breach cost Ahold Delhaize about €200M in lost sales. [2]
The UK's NCSC judged it highly likely that AI-powered malware will be used in a ransomware attack on UK critical infrastructure by 2028. India's CERT-In warned that a group called CRAFTY FALCON is using publicly available automated intrusion tools against lenders, fintechs and payment firms. [6] [7]
What the sources show: disclosures and exploit attempts rose sharply, and agencies issued warnings specific to AI.
Read the full assessment
What it may mean, though not yet proven: AI is shortening the time from discovery to exploitation, so patching, disclosure coordination and developer-device controls may need to move faster.
Executive brief
AI is now finding serious bugs faster than companies can patch them. Reuters reported that about 37,000 software vulnerabilities were disclosed in Q3 2026, roughly three times the figure a year earlier (Insurance Journal/Reuters). The clearest example came in October: AI-assisted researchers found critical PHP flaws, details leaked before the fix shipped, and Cloudflare counted millions of exploit attempts (TIME). Regulators in the UK and India have since issued AI-specific warnings. Only the feed excerpt of the WSJ newsletter was available. It also flags an FTC case against Amazon and China's control over AI, neither of which is covered here.
What changed and event timeline
- Early Aug 2026
An AI agent finds a PHP flaw
A Wiz team was using an AI agent to work on a Drupal weakness when it found an underlying PHP bug. Calif, working with OpenAI, found a related flaw separately ().
Emergency PHP patches
PHP shipped fixes for CVE-2026-109618/109619 in versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Amazon reportedly saw suspicious activity aimed at the flaws on Oct 17 ().
Record $6.3B Binance theft
A compromised DevOps laptop led to illicit withdrawals. TRM Labs and Chainalysis attributed the attack to North Korea's TraderTraitor group. No AI involvement is claimed ().
NCSC annual review
The UK's NCSC judged it highly likely that AI-powered malware will be used in a ransomware attack on UK critical infrastructure by 2028 ().
CERT-In critical advisory
Advisory CIAD-2026-0058 warns that a group called CRAFTY FALCON is using AI tooling to attack Indian lenders, fintechs and payment firms ().
Capabilities and access
- PHP discovery: TIME names no specific model. It says Wiz used "an AI agent" and Calif worked with OpenAI, which gave defenders early access (TIME). A lower-trust site credits "GPT 5.6-Sol" (shattered.io).
- Attacker tooling: CERT-In lists ARTEX, PentAGI, Cairn and CyberStrikeAI. It describes them as publicly available tools that automate intrusion steps with little human input (Open Chronicle).
Read the full section
- PHP discovery: TIME names no specific model. It says Wiz used "an AI agent" and Calif worked with OpenAI, which gave defenders early access (TIME). A lower-trust site credits "GPT 5.6-Sol" (shattered.io). TIME and National CIO Review don't confirm that name.
- Attacker tooling: CERT-In lists ARTEX, PentAGI, Cairn and CyberStrikeAI. It describes them as publicly available tools that automate intrusion steps with little human input (Open Chronicle).
Technical analysis for researchers and developers
- Vulnerability class: both PHP flaws involve memory handling and could allow code execution on servers without authentication (National CIO Review).
- Agentic exploitation: the Wiz agent found the PHP flaw while trying to exploit a different target (Drupal).
- Cost: Claroty said it recreated an attack for about $6 in 28 minutes (Insurance Journal/Reuters).
Read the full section
- Vulnerability class: both PHP flaws involve memory handling and could allow code execution on servers without authentication (National CIO Review). TIME reports the flaws were long-standing but hard to exploit.
- Agentic exploitation: the Wiz agent found the PHP flaw while trying to exploit a different target (Drupal). In other words, the agent was chaining exploits, not just scanning code. Researchers said work that used to take months now takes hours to days (TIME).
- Cost: Claroty said it recreated an attack for about $6 in 28 minutes (Insurance Journal/Reuters).
- Reproducibility: none of the sources describe a methodology, benchmark or model configuration.
Claims and evidence
- 37,000 Q3 vulnerabilities, about 3× year over year: reported by Reuters; the underlying dataset isn't named (Insurance Journal).
- Exploit surge: Cloudflare's own telemetry, as reported by TIME.
- Exposure: Censys counted nearly 55 million web properties running PHP (National CIO Review). That counts PHP sites, not confirmed vulnerable ones.
Read the full section
- 37,000 Q3 vulnerabilities, about 3× year over year: reported by Reuters; the underlying dataset isn't named (Insurance Journal).
- Exploit surge: Cloudflare's own telemetry, as reported by TIME. Attempts went from thousands before the announcement, to tens of thousands within 45 minutes, to hundreds of thousands within two hours, and eventually millions (TIME).
- Exposure: Censys counted nearly 55 million web properties running PHP (National CIO Review). That counts PHP sites, not confirmed vulnerable ones.
- Binance funds frozen: Binance says about $3B (ITV). BleepingComputer reports about $2.7B frozen by stablecoin issuers, plus more frozen on Arbitrum.
- CRAFTY FALCON's use of AI: this comes from a government advisory. The reviewed sources contain no independent corroboration.
Context and prior work
- In December 2025, security researcher Rachel Tobac predicted that malicious hackers would get a big AI boost in 2026 (Axios).
- At $6.36B, the Binance theft is about $4.86B larger than the February 2025 Bybit hack, also attributed to North Korea (BleepingComputer).
- The NCSC says China is best positioned to use AI against UK critical infrastructure (Insurance Journal).
Read the full section
- In December 2025, security researcher Rachel Tobac predicted that malicious hackers would get a big AI boost in 2026 (Axios).
- At $6.36B, the Binance theft is about $4.86B larger than the February 2025 Bybit hack, also attributed to North Korea (BleepingComputer).
- The NCSC says China is best positioned to use AI against UK critical infrastructure (Insurance Journal).
Limitations, safety and contested findings
- Disclosure failure: Calif mistakenly posted an exploit video hours before the patch was available.
- Post-mortem dispute: a post-mortem reportedly found that a small group of maintainers had known about the bug for about two years, which complicates the claim that AI discovered it (Tom's Hardware).
- Attribution limits: the year's biggest incident, Binance, is not linked to AI. The headline vulnerability counts don't separate AI-found bugs from others.
Read the full section
- Disclosure failure: Calif mistakenly posted an exploit video hours before the patch was available. That forced PHP to release a day early, and exploitation started before the fix shipped (TIME).
- Post-mortem dispute: a post-mortem reportedly found that a small group of maintainers had known about the bug for about two years, which complicates the claim that AI discovered it (Tom's Hardware).
- Attribution limits: the year's biggest incident, Binance, is not linked to AI. The headline vulnerability counts don't separate AI-found bugs from others.
Business and practitioner implications
- Patch speed: Ionos says serious flaws now arrive weekly or every few days.
- Defensive AI: Booking.com, the Port of Rotterdam and Siemens are deploying or testing AI defenses.
- Fintech controls: CERT-In recommends round-the-clock monitoring, limits on payment API transactions, separating development from production, and confirming that each fix actually closed the hole (Open Chronicle).
Read the full section
- Patch speed: Ionos says serious flaws now arrive weekly or every few days. Ahold Delhaize reported about €200M in lost sales after an October breach (Insurance Journal/Reuters).
- Defensive AI: Booking.com, the Port of Rotterdam and Siemens are deploying or testing AI defenses. McKinsey projects the security market will grow from $223B in 2025 to $346B in 2029 (same source).
- Fintech controls: CERT-In recommends round-the-clock monitoring, limits on payment API transactions, separating development from production, and confirming that each fix actually closed the hole (Open Chronicle).
- Insider-access risk: the Binance breach started from one engineer's laptop, so access to developer and DevOps devices needs tighter control.
The source trail.
Sources (11)
The Morning Risk Report: Companies Face Cyber Pain as Hackers Press AI Advantage
Feed excerpt only.
wsj.com