Wikipedia: PHP2Shell
Copy:
PHP2Shell
PHP2Shell is a set of two high-impact vulnerabilities (CVE-2026-109618 and CVE-2026-109619) in PHP versions 5.4–8.5 that allow remote code execution (RCE). Their existence was publicly announced on 20 October 2026.[2] An emergency patch for versions 8.2, 8.3, 8.4 and 8.5 and details about the vulnerabilities were released the following day.[3]
Discovery
PHP versions as early as 5.4 (released in 2012)[4] are affected by these vulnerabilities, and while concerns around memory safety bugs had been raised decades before, no action was taken at the time because it was not thought that developing a working exploit could be feasible.[5][6] A PHP core developer discovered the out-of-memory error behind one of the vulnerabilities during an audit in 2024, but the bug was never fixed; the development team could not reach consensus on how to do so and a specific exploit had not yet been discovered, so the issue was not prioritized.[7][8]
The vulnerabilities were independently discovered by two parties: Ryan Austin, a researcher at the California-based[9] security firm Calif, as part of the OpenAI "Patch the Planet" project in August 2026; and the Israeli–American cloud security company Wiz, a subsidiary of Alphabet Inc. and a part of Google Cloud. Austin used OpenAI's GPT-5.6 Sol model to assist with finding the vulnerability.[10] Wiz researchers used the large language models Gemini 4 Argon and GPT-5.6 Sol, as well as their own tool, Wiz Atlas, to help with their research.[11]
The vulnerability was rated 9.2 (critical) out of 10 under CVSS v4.0 by PHP developers,[12] while security firm watchTowr rated it a full 10.0.[13]
Patches
The maintainers of PHP released patched versions of PHP 8.2, 8.3, 8.4, and 8.5 (8.2.35, 8.3.36, 8.4.27, and 8.5.12).[3] Source code patches were released for version 7.4, 8.0, and 8.1, but no official updated binaries for those versions were released as they are considered end of life.[3] Temporary mitigations were also published, but the maintainers said there were no complete mitigations and that updating was the only safe choice.[3] Both vulnerabilities were published through GitHub security advisories on 21 October 2026 at 13:34 UTC,[14][15] 26 minutes before the announced release time, with the PHP website updated shortly thereafter.[citation needed]
The PHP maintainers announced the update early and before release, after news of the vulnerabilities leaked on 20 October on the Hacker News forums[16], in a post on X by cybersecurity company watchTowr labs, and in a proof-of-concept exploit video mistakenly released by Calif.[17][unreliable source?][18][7] The early notice on PHP.net encouraged users to prepare and update as soon as the patch was released.[3] One security firm, Sansec, expected "mass attacks to start within 15 minutes of the release".[19] Proof-of-concept exploits, write-ups, and partial mitigations started circulating on the Internet before the end of the embargo, after development platform Chainguard released patched versions of the software early; this allowed attackers to determine what code had been changed.[13][20][7]
Amazon Threat Intelligence identified a PHP2Shell exploitation attempt dating back to 17 October 2026, in MadPot telemetry, four days before public disclosure.[21] Cloudflare reported signs of early exploitation prior to public patch release, and deployed temporary web application firewall (WAF) rules as an attempted stopgap mitigation against the vulnerability.[22][23][24] Security firm watchTowr said it had observed exploitation in the wild by 21 October through its honeypot network,[25] while the US Cybersecurity and Infrastructure Security Agency (CISA) said it had not observed exploitation as of the same day.[26]
Impact
Approximately 69% of websites use PHP, including those that use content management systems such as Drupal, Joomla, and WordPress, so the potential impacts were widespread.[26][27] PHP is also bundled with vendor appliances, web host control panels such as cPanel and Plesk, and some Windows applications.[28]
Following advice from CERT-EU and the British National Cyber Security Centre, many organizations across Europe took their websites offline as a precaution while the patches were pending, including the Dutch government portal Overheid.nl, the Danish national school communication platform, all Finnish government public-facing web services, and all websites of the French Interior Ministry.[11][27][29][30] The US National Security Agency recommended that websites go offline if unable to update immediately.[5][31] The UK National Cyber Security Centre released a similar warning, urging organizations to be wary of backdoors.[32] The Japanese National Cybersecurity Office advised PHP users to immediately update their systems.[33]
The Wikimedia Foundation shut down many of their supporting services, and forum sites 4chan and Something Awful went offline.[26][34][5] Parts of the Japanese internet also went offline, including games such as Umamusume: Pretty Derby, Granblue Fantasy, Princess Connect! Re:Dive, and many online commerce sites.[33][34] The Internet Archive briefly shut down to fix the issue.[35]
Analysis
Researchers at Wiz stated that the issue was on par with Heartbleed; a 2014 vulnerability that affected an encryption library used by about 2 thirds of the web at the time.[36] Commentary in the Guardian noted that the PHP Foundation, which maintains the PHP language, only had 13 full- and part-time employees, and relied on financial contributions, whereas AI research, which was credited with helping discover the vulnerability, received "billions" in investment.[37]
---
[AINews]
Top Story: PHP2Shell—chronology, core claims, and the complications behind the emergency response
What happened
Researchers disclosed two PHP-engine memory-corruption vulnerabilities, with demonstrated unauthenticated exploitation against major CMSs and an unusually broad coordinated response.
The researchers named the vulnerabilities PHP2Shell and identified them as CVE-2026-109618 and CVE-2026-109619. Their technical description places the defects in PHP’s internal handling of out-of-memory conditions and fatal errors—not in a WordPress plugin or an application’s authentication logic. They report demonstrations of full remote code execution and “Heartbleed-style” memory disclosure. @sagitz_
The disclosed demonstrations included pre-authentication RCE against WordPress, Drupal, and Joomla. The researcher’s announcement separately claims remote memory disclosure against any PHP application. These are researcher-reported exploit results; they should not be silently expanded into a claim that every deployment has an identical, immediately exploitable RCE path. @sagitz_
OpenAI says Calif researchers used its models to discover the flaws, while secondary reporting says Wiz independently found the same bugs. Posts identify GPT-5.6 Sol on the Calif/OpenAI side; a Wiz-affiliated account credits Gemini 4 Argon with discovery, exploit-chain development, detector creation, and patch validation. This is a multi-party discovery story, not evidence that one vendor alone found and fixed the issue. @TheRundownAI, @steipete, @ralucaadapopa
OpenAI describes approximately two months of coordinated remediation before disclosure. It says the patch required a major PHP-engine change and coordination among maintainers, researchers, security providers, and governments. A participating researcher confirms work with the PHP Foundation to support and harden the patch. @OpenAI, @itszn13
Cloudflare deployed initial WAF mitigations before the scheduled patch release. Its CTO said rules covered known exploit vectors, were available to all customers including the free tier, and were being updated as early attacks emerged. He explicitly called the rules a stopgap and directed operators to patch at 14:00 UTC. @dok2001
The supplied NSA account advised preparation for an immediate update—and taking systems offline if timely mitigation was not possible. Later OpenAI messaging said upgrades were available; secondary reporting says details leaked before the patch. The supplied material does not establish precisely what leaked, when, or whether the release schedule changed. @NSACyber, @OpenAI, @TheRundownAI
Reconstructing the chronology
1. Discovery and private coordination: approximately two months before public disclosure.
OpenAI places the work under Patch the Planet and says it led a coalition preparing fixes and protections. Its security researcher Eric Wallace adds that internal teams worked on patching and detection rather than merely disclosing the bug, with confidential advance notification to governments and outside agencies. Greg Brockman reiterates the two-month coordination period. These are participant accounts, not an independently audited incident timeline. @Eric_Wallace_, @gdb
The distinction between program names matters: the posts associate the vulnerability with Patch the Planet, while researchers credit Daybreak model access for enabling security work. Those are related roles—remediation initiative versus access mechanism—not interchangeable descriptions of the same program. @OpenAI, @S1r1u5_
2. Advance public warnings and a scheduled patch window.
Japanese-language posts warned operators to prepare for a 23:00 JST, October 21 release, consistent with Cloudflare’s 14:00 UTC announcement. One correctly emphasized that a PHP-core vulnerability cannot be fixed merely by waiting for plugin developers. Another framed the event as a “15-minute” race after patch publication, but that number was a prediction—not a measured time to weaponization. @ds_nakajima, @expokun
3. Pre-release filtering and observed attack activity.
Cloudflare’s post is the strongest firsthand operational evidence in the supplied record: mitigations were already deployed, known vectors were covered, and “early attacks” were appearing. It does not specify successful compromises, request counts, victim counts, payload provenance, or whether the traffic represented original exploit development versus reuse of leaked information. @dok2001
4. Research disclosure, government warning, and patch availability.
The researcher’s public thread describes the exploit demonstrations and error-handling root cause. The NSA post calls for urgent preparation and conditional shutdown. OpenAI subsequently says every PHP site can upgrade safely. The supplied set contains no direct PHP Foundation release advisory, exact fixed-version matrix, distribution package identifiers, or patch commit. That missing primary-source layer prevents a reliable “upgrade to version X” instruction. @sagitz_, @NSACyber, @OpenAI
5. Alleged premature exposure and widespread service disruption.
The Rundown reports that details leaked early. A later viral compilation adds approximately 1.5 million servers, attributed to Censys, running versions without a fix; reported shutdowns involving Dutch institutions, government sites elsewhere in Europe, Wikipedia, Archive.org, Asana, GOG, and others; and a 05:55 warning from Japan’s prime minister. None of those individual claims is corroborated by its underlying advisory, measurement, or status page in the supplied material. Nor does the record establish whether each outage was a precaution, a patching interruption, an unrelated failure, or a compromise. @TheRundownAI, @fiveoutofnine
SemiAnalysis separately reports that Japan’s prime minister urged PHP operators to patch immediately or temporarily disconnect if they could not. Its account contextualizes that response with an existing Japanese breach wave, rather than establishing that PHP2Shell caused those earlier incidents. @SemiAnalysis_
6. Exposure-checking assistance followed disclosure.
OpenAI published a repository containing a Codex skill intended to guide operators through checking whether their sites were affected. That is an assistance workflow, not a substitute for a vendor advisory, package verification, or evidence that every vulnerable runtime has been replaced. @OpenAI
AI’s role: credible evidence, competing credit, and unresolved causality
The affirmative case is stronger than a benchmark-only claim. Researchers describe working exploit chains, coordinated patch engineering, detectors, and patch validation. A participating researcher calls PHP2Shell the most consequential vulnerability they have helped discover. OpenAI characterizes it as an example of the “Defender’s Window”: advanced models finding a flaw before attackers, giving defenders time to prepare. Importantly, its wording is “to our knowledge”—not proof that no attacker previously knew the vulnerabilities. @itszn13, @OpenAI
Credit is distributed across humans, models, and infrastructure providers. Calif/OpenAI and Wiz/Gemini appear as separate discovery accounts; PHP maintainers supplied the upstream remediation path; Cloudflare supplied pre-release filtering. Google also advertises Wiz testing in which Gemini 4 Argon found more than twice as many vulnerabilities as baseline models, but the supplied promotional post does not expose the dataset, baselines, budgets, or false-positive accounting. It is not an independent comparative evaluation of PHP2Shell discovery performance. @TheRundownAI, @ralucaadapopa, @Google
Researcher-access policy draws both praise and criticism. One researcher credits OpenAI’s Daybreak access with a portfolio of findings spanning PHP, GitLab, Chrome, WordPress, and other targets. Another sharply contrasts OpenAI’s vetted access with Anthropic’s restrictions; the claim that Anthropic contributed “zero value” is that author’s sweeping opinion, not established by the PHP record. A further post criticizes restrictions on Astra access and argues for giving 1,000 vetted researchers the best available models with intensive monitoring, rather than a rigid red/blue distinction. @S1r1u5_, @migtissera, @S1r1u5_
The countervailing safety rationale appears in reporting that Anthropic delayed broad access to Mythos-class models for four months to strengthen cyber safeguards. That provides context for restricted access, but does not settle whether its implementation helped or hindered this particular remediation effort. @kimmonismus
A burst of disclosures is not, by itself, evidence of a burst of AI-caused compromises. The supplied feed mixes defensive discovery, supply-chain attacks, credential theft, precautionary shutdowns, and unrelated vulnerabilities. Ethan Mollick interprets the moment as attack and defense accelerating beyond human response speeds; John Allard describes a capability “phase transition” while acknowledging that many recent findings came from white-hat programs. Both are interpretations of the pattern, not incident-level causal attribution. @emollick, @john__allard
Operational complications: patching, shutdowns, and the temptation to “fix everything”
The emergency is as much distribution and ownership as discovery. Operators need to locate PHP across managed hosting, containers, appliances, background workers, and old applications; obtain appropriate packages; deploy them; restart or replace affected processes; and verify the running build. One operator describes maintaining numerous PHP/WordPress sites and being exhausted by testing and updates. Another concludes that patch distribution is poorly coordinated. @ostrisai, @jd_pressman
WAF deployment reduces immediate exposure but creates a false-confidence risk. Cloudflare’s own statement is narrower than the later advice to put all internet-facing services behind it: known vectors were covered, protections were evolving, and patching remained mandatory. Proxied coverage, alternate ingress paths, and the origin’s exposure must be checked rather than assumed. @dok2001, @threepointone
Shutdown guidance is conditional, not evidence that the entire internet should disconnect. The supplied NSA post recommends taking systems offline when immediate updating is infeasible until full mitigation is possible. For a specific operator, the relevant questions are affected-version status, exposure, available mitigations, and the consequences of continued operation—not the engagement score of an alarming post. @NSACyber
“Use the RCE to auto-patch every server” is a proposal, not an authorized remediation plan. Marc Köhlbrugge raises the idea; J. D. Pressman suggests a presidential pardon for a lab or white-hat team attempting it. Unsolicited exploitation still risks unauthorized access, outages, incorrect fixes, destruction of evidence, and cross-jurisdictional liability. A speculative pardon does not provide universal authorization or eliminate civil and foreign legal exposure. @marckohlbrugge, @jd_pressman
Update velocity has a supply-chain trade-off. One widely shared post captures the dilemma: updating constantly can expose users to compromised packages, while delaying updates leaves known vulnerabilities open. Pressman argues that endless patching is unsustainable unless the rate of new vulnerabilities and risky releases falls; he also notes that receiving no patch and discovering theft is worse than receiving too many updates. @xeophon, @jd_pressman, @jd_pressman
The longer-term alternatives remain open questions. Sarah Guo reports CVEs increasing from roughly 4,000 per month to 15,000 since April, and asks about formal verification, safer distributions, and organizations without red teams. The counts are not independently verified here, and disclosure volume is not a direct measure of successful exploitation. Other posts ask what a full-codebase security sweep would cost and recast the challenge as hardening mundane infrastructure rather than debating only abstract catastrophe. @saranormous, @saranormous, @teortaxesTex, @mimi10v3
For an incident team, the defensible sequence is to obtain the authoritative PHP and distribution advisories, inventory exposed runtimes, apply temporary protections where justified, deploy authenticated fixed packages, verify running processes, and investigate evidence of exploitation separately. Updating closes a vulnerability; it does not establish that credentials or data were never exposed.
Public reactions: practical caution versus unsupported escalation
The precautionary perspective stresses that interconnected services are fragile and remediation can be costly. Joshua Achiam compares the moment to March 2020, speculates about internet lockdowns, recommends backing up valued internet material, and suggests printing important records. Pressman says remediation costs exceeded his expectations. These are preparedness judgments—not evidence of impending systemic collapse. Paper copies may help continuity but are not inherently tamperproof logs. @jachiam0, @jachiam0, @jachiam0, @jd_pressman
The alarm-skeptical perspective asks for specificity. Alex Johnson argues that warnings about bank accounts being arbitrarily emptied obscure the distinction between unauthorized transfers and scams in which customers authorize payment. He invokes U.S. Electronic Funds Transfer Act protections with timely reporting; those protections are jurisdiction-, account-, and circumstance-dependent, not a universal guarantee. Kelsey Piper separately warns about the psychological cost of living under constant annihilation framing. @AlexH_Johnson, @KelseyTuoc
Broad personal-security advice needs to be separated from the PHP fix. Password managers, unique credentials, strong MFA, patching, carrier port-out protections, and U.S.-specific credit or tax identity safeguards can reduce downstream harm. But enabling 2FA does not prevent exploitation of an unauthenticated PHP-engine bug. Likewise, viral advice to avoid all online payments the next day is not supported by a demonstrated scope assessment. @peterwildeford, @StillMare_
The broader threat context is real but not interchangeable with PHP2Shell. SemiAnalysis reports 18 Japanese companies disclosing breaches since September 25, including exposures of approximately 6.6 million Times Car accounts, 11.18 million Burger King Japan app records, and potentially 22.5 million Sourcenext records. It attributes findings about AI-assisted scanning and API attacks to JPCERT/CC. Another post quotes an August 19 advisory concerning AI-assisted tooling against internet-exposed Siemens PLCs. These are separate reported incidents and advisories; neither establishes PHP2Shell as their entry point. @SemiAnalysis_, @AndrewCurran_
The key unresolved evidence gaps are concrete: exact affected and fixed versions; CVSS vectors; deployment prerequisites; the independent discovery timeline; what leaked before release; successful exploitation versus scanning; validated exposure counts; and the cause and duration of each reported shutdown. The supplied viral incident compilation cannot answer those questions merely by placing PHP, GitLab, LiteLLM, Railway, certificate-authority failures, and government warnings in the same narrative. @fiveoutofnine
---
Email của Dương Ngọc Thái gửi nhà báo Vũ Kim Hạnh:





Không có nhận xét nào:
Đăng nhận xét