--- title: "PHP2Shell Exploited by AI Agents: 404 Hosts, What to Do" url: https://playciso.com/blog/php2shell-ai-agent-credential-harvesting updated: 2026-11-21 section: Blog publisher: PlayCISO --- # PHP2Shell Exploited by AI Agents: 404 Hosts, What to Do > SOCRadar: an AI-agent-assisted operator used PHP2Shell (CVE-2026-109618/109619) on 404 hosts and took cloud keys. Fixed versions, what to do. Source: https://playciso.com/blog/php2shell-ai-agent-credential-harvesting · Published: 2026-11-19 · Publisher: PlayCISO (https://playciso.com) Primary source: https://socradar.io/blog/php2shell-ai-agents-credential-harvesting/ --- **An operator used AI agents to exploit PHP2Shell at scale, and SOCRadar says 404 servers were compromised.** In a report published on 18 November 2026, [SOCRadar](https://socradar.io/blog/php2shell-ai-agents-credential-harvesting/) describes a campaign against the two critical PHP flaws CVE-2026-109618 and CVE-2026-109619 (PHP2Shell) that began on **21 October 2026**, the day PHP shipped fixes and public proofs of concept appeared. It counts **595 targeted IPs in 57 countries**, **404 compromised hosts** and about **290,000 exfiltrated files (9.5 GB)**, much of it cloud keys and application secrets. SOCRadar learned this from the operator's own servers, which were left open. The figures are SOCRadar's; no victim has confirmed them, and the operator is not identified. The fix is **PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35**, but a server exposed before it was patched may still be compromised. **Update, 19 November 2026:** PHP shipped the announced security releases, **8.5.13, 8.4.28, 8.3.37 and 8.2.36** (plus 8.6.0). They fix, among others, [phpAfterShell (CVE-2026-126714)](/blog/phpaftershell-cve-2026-126714), a critical session-file flaw that depthfirst says bypasses the October PHP2Shell fix, and CVE-2026-129481, which PHP describes as an incomplete fix for CVE-2026-109619. The October releases below stop PHP2Shell but are affected by both, so update again. **Who this is for:** CISOs, vulnerability-management and platform leads, and anyone who runs PHP on internet-facing servers, including Nextcloud and custom PHP applications on cloud instances. *Based on SOCRadar's report, php.net, the CVE records, NVD and the CISA KEV catalog, all read on 19 November 2026. PlayCISO is not affiliated with SOCRadar or the PHP project. We publish no exploit details and no operator indicators; SOCRadar's post has the indicator list.* Confirmed vs not yet known **PHP confirms:** two critical flaws that can allow unauthenticated remote code execution, fixed on 21 October 2026 in 8.5.12, 8.4.27, 8.3.36 and 8.2.35, each CVSS v4.0 **9.2**. PHP says every installation serving web requests is at risk (PHP-FPM, Apache mod_php, php-cgi, FrankenPHP, LiteSpeed), with the highest risk under mod_php, and that there are no complete mitigations. **CISA confirms exploitation:** both CVEs were added to the Known Exploited Vulnerabilities catalog on 26 October 2026 (due 29 October, forensic triage required), and CISA's enrichment of the CVE records says "Exploitation: active". **Reported by SOCRadar (not confirmed by any victim):** one operator's campaign, reconstructed from its exposed servers: 595 targeted IPs in 57 countries, 404 compromised hosts, about 290,000 files (9.5 GB) exfiltrated by 3 November, an AI agent queue of 39 workers backed by two DeepSeek V4.1 endpoints, and about 3.3 million user records taken from an unauthenticated Redis instance. **Not established:** who the operator is (SOCRadar's confidence in identity is low), how much of the campaign ran autonomously (SOCRadar says it could not confirm it), which organisations were hit (SOCRadar names none), and whether any cryptocurrency was stolen (SOCRadar says no theft was confirmed). ## Affected and fixed PHP versions | Branch | Affected (per PHP's CVE records) | Fixes PHP2Shell (21 Oct 2026) | Take now (19 Nov 2026) | | --- | --- | --- | --- | | PHP 8.5 | 8.5.x before 8.5.12 | 8.5.12 | 8.5.13 | | PHP 8.4 | 8.4.x before 8.4.27 | 8.4.27 | 8.4.28 | | PHP 8.3 | 8.3.x before 8.3.36 | 8.3.36 | 8.3.37 | | PHP 8.2 | 8.2.x before 8.2.35 | 8.2.35 | 8.2.36 (8.2 reaches end of life on 31 December 2026) | | PHP 8.1, 8.0, 7.4 | No longer supported by the PHP project | Minimal source backports from PHP | Your distribution or LTS vendor | - CVE-2026-109618 ([GHSA-w8hw-w9fx-j79w](https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w)), "RCE via OOM corruption": PHP says out-of-memory handling can interrupt it while internal request data is only partly initialised, and request cleanup may then treat that data as valid. - CVE-2026-109619 ([GHSA-6wmf-4rv5-p3rp](https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp)), "RCE via stale global state": PHP says request-owned internal state can survive after its memory is freed, and a later request handled by the same worker may use it. - Severity: both CVSS v4.0 9.2 from PHP (network, high attack complexity, attack requirements present, no privileges, no user interaction). CISA's enrichment rates them "Automatable: no", "Technical impact: total". - Distribution packages: PHP says Linux distributions and LTS vendors received the patches in advance, so a distro package may carry the fix under a different version string. Check your vendor's advisory. - The 19 November releases. PHP pre-announced them on 18 November with highest severity CRITICAL and shipped 8.5.13, 8.4.28, 8.3.37 and 8.2.36 on 19 November. Per the ChangeLog and PHP's advisories they fix [CVE-2026-126714](https://github.com/php/php-src/security/advisories/GHSA-fm2m-fjf5-v3cp) (phpAfterShell, CVSS v4.0 9.2; see [our post](/blog/phpaftershell-cve-2026-126714)), CVE-2026-115560 (the session serializer issue behind the 23 October advice) and [CVE-2026-129481](https://github.com/php/php-src/security/advisories/GHSA-g3vh-838c-65v7) (an incomplete fix for CVE-2026-109619, 8.3 to 8.5), among other flaws. Neither of the two session CVEs is in KEV as of 19 November. Our [PHP 8.5.13 security release post](/blog/php-security-release-8-5-13-cve-2026-129289) covers CVE-2026-115560, the PHP-FPM request smuggling flaw CVE-2026-129289 and CVE-2026-129474. If you cannot update at once, PHP lists partial measures that "do not make an installation safe": for CVE-2026-109618, `memory_limit=-1` set where applications cannot override it, a low `post_max_size`, and request-size, field-count, part-count and header-count limits at the web server or proxy; for CVE-2026-109619, `session.upload_progress.enabled=0` with every worker restarted, and on PHP 8.5 `fatal_error_backtraces=0`. Treat them as a bridge to the update, not a substitute. ## What SOCRadar found SOCRadar's Threat Research Unit says it found the operator's command-and-control (C2) server serving its home directory and loot store over open, unauthenticated HTTP, and on 8 November a second open directory on the operator's AI lab server with an 18 GB working tree. Everything in this section is **SOCRadar's account**. - Disclosure-day start. SOCRadar says the operator wrote its first PHP2Shell exploit script on 21 October, the day public proofs of concept were released, and adapted it within days to default PHP/Apache configurations and to individual PHP builds. It audited about 13 public PHP2Shell repositories on 23 October to fold in per-build details. - How it got in, at a high level. SOCRadar describes a large multipart request that exhausts PHP's default memory limit, followed by a second request on the same keep-alive connection that leaks memory and then runs a command. It says the operator's version works against default settings, under mod_php and FrankenPHP, and failed under PHP-FPM; PHP itself says FPM is also at risk, so do not deprioritise FPM hosts. We do not repeat the technique. - Target sourcing. Targets came from internet search data: a pool of 68,198 PHP/Apache hosts, shortlisted to about 5,042, with CDN- and WAF-fronted sites filtered out. Hosts already on the fixed release produced no leak and no shell in the operator's own records. - Scale. 595 unique IPs in 57 countries targeted; 404 hosts compromised; interactive shells documented on 217 hosts but root on only 12, as most access ran as the unprivileged www-data user inside containers. - Who was hit. SOCRadar says targeting followed PHP/Apache exposure, not industry. Of 452 collection jobs, 200 were custom PHP applications and 163 Nextcloud; at least 68% ran in Docker or Podman containers, and Amazon EC2 was the most common host (112 jobs). Where an organisation could be identified (about 120 jobs) the sectors were marketing and media analytics (88 jobs, one AWS-hosted estate), education, IT services and hosting, financial services and payments, healthcare, energy and water, and software. It names no victims. - Persistence. Operator SSH keys, reverse-shell beacons re-run every 60 seconds from cron, token-protected PHP webshells, application-level backdoors (a Nextcloud app and a Laravel route-cache backdoor) and a Dropbear SSH server bound to localhost with a reverse tunnel. - What was taken. About 290,000 files (9.5 GB) across 481 job directories as of 3 November: cloud API keys and role credentials read from the instance metadata service, 191 application .env files across 144 jobs, database dumps, OAuth tokens, Git repositories, session files, private keys and one mailbox export. - The money track. A separate, agent-driven effort against payment platforms reached account takeover and withdrawal capability. SOCRadar says no theft was confirmed: a withdrawal of about 456,900 USDC stayed pending, a second platform was unreachable, and encrypted wallets were not cracked. The completed impact was about 3.3 million user and session records (about 1 GB) taken from an unauthenticated Redis instance. SOCRadar's own counts differ slightly by method: its headline is 404 compromised hosts across both exposed directories, while the first set of collection jobs alone establishes compromise on about 393 unique hosts. ## How the AI orchestration worked, and what is not established SOCRadar says the operator ran a self-built job queue it calls **maq** on a dedicated server. The queue dispatched eligible targets to **39 parallel agent workers** (patched workers of a tool SOCRadar names Oh My Pi, or omp), load-balanced across **two self-hosted DeepSeek V4.1 model endpoints** on rented GPU machines. A monitor polled the C2 database every 15 seconds, and 194 C2 entries reference an agent job. The workers verified target stacks, triggered memory leaks, worked out per-build offsets, created C2 records and ran the data collector. One planning file, GOAL.md, told an agent to keep working "autonomously until a real payout is executed", aiming at more than $50,000 in USDT. It failed. SOCRadar also notes that the worker's environment carried a variable set by Anthropic's Claude Code command-line tool, which it reads as showing the agent harness was Claude Code or compatible with it, "not which model answered"; the queue configuration points the workers at the DeepSeek endpoints. **What is not established:** - The degree of autonomy. SOCRadar says the snapshot "cannot clarify" whether actions came from autonomous agents or from the operator working in the same sessions, so "the true degree of automation remains unconfirmed". Its conclusion: "human-directed and AI-augmented", and "no chain completed without manual oversight". - Novel capability. SOCRadar rates the technical complexity "moderate" and the tradecraft "average": an adapted public exploit plus custom orchestration, with no new memory-corruption research. - Identity. No known group is linked. SOCRadar assesses with moderate confidence a financially motivated operator working like an initial-access broker, and says confidence in identity is low. The lesson for defenders is speed, not science fiction: an average operator with an agent queue went from public proof of concept to hundreds of compromised servers in days. An AI worker even ran the collector on the operator's own lab and uploaded the result to the open loot store, which is part of how SOCRadar saw so much. Other firms have published separate reports about PHP2Shell activity and different actors. They are separate observations and are not counted in any figure above. ## PAN-OS firewalls: PHP2Shell + Fragnesia (added 20 November 2026) A separate campaign hit Palo Alto Networks firewalls, whose management interface runs PHP. Palo Alto Networks' [advisory CVE-2026-109618](https://security.paloaltonetworks.com/CVE-2026-109618) says PHP2Shell affects PAN-OS on PA-Series and VM-Series firewalls and on Panorama and has been exploited against management interfaces exposed to the internet; fixed hotfixes were out on every supported branch by 29 October. [Dream Research](https://dreamgroup.com/blog/chained-php2shell-exploit-roots-25-servers) reported that one attacker chained PHP2Shell with Fragnesia (CVE-2026-46300), a Linux kernel privilege escalation flaw that CISA added to KEV on 28 October, to get root on 24 firewalls in the United States, Japan, Taiwan and South Korea and one internal management server; Unit 42 says it verified the tooling and results. Neither SOCRadar nor Dream links the two campaigns, so the figures are not combined. Fixed PAN-OS versions and what to do: [PAN-OS PHP2Shell + Fragnesia: 25 systems rooted](/blog/pan-os-php2shell-fragnesia-root). ## Timeline | Date | Event | Source | | --- | --- | --- | | 20 Oct 2026 | PHP pre-announces a security release for two critical flaws, both CVSS v4.0 9.2 | php.net | | 21 Oct 2026 | PHP releases 8.5.12, 8.4.27, 8.3.36 and 8.2.35 with an advisory for CVE-2026-109618 and CVE-2026-109619; CVE records published | php.net, CVE record | | 21 Oct 2026 | Public proofs of concept appear; the operator writes its first exploit script and the C2 catalogue starts recording targets (21-25 Oct) | SOCRadar | | 22 Oct 2026 | The operator's first two hand-run waves produce most of the recorded shells | SOCRadar | | 23 Oct 2026 | PHP informational advisory: set session.upload_progress.enabled=0 against a further, unlikely RCE vector (CVE-2026-115560) | php.net | | 23 Oct 2026 | The operator audits about 13 public PHP2Shell repositories for per-build details | SOCRadar | | 26 Oct 2026 | CISA adds both CVEs to the KEV catalog (due 29 Oct, forensic triage required) | CISA KEV | | 29-30 Oct 2026 | A second operator exploit and SSH source appears across 70 jobs | SOCRadar | | 3 Nov 2026 | Loot store snapshot: about 290,000 files, 9.5 GB, 481 job directories | SOCRadar | | 8 Nov 2026 | A second open directory is found on the operator's AI lab server (18 GB working tree) | SOCRadar | | 18 Nov 2026 | SOCRadar publishes its report | SOCRadar | | 18 Nov 2026 | PHP announces further security releases for 19 November, highest severity CRITICAL | php.net | | 19 Nov 2026 | PHP releases 8.5.13, 8.4.28, 8.3.37 and 8.2.36, fixing phpAfterShell (CVE-2026-126714) and CVE-2026-129481, an incomplete fix for CVE-2026-109619 | php.net, PHP advisory | ## What security leaders should do this week - Find every PHP that serves web traffic and patch it. Move to 8.5.13, 8.4.28, 8.3.37 or 8.2.36, the 19 November releases (the October 8.5.12, 8.4.27, 8.3.36 and 8.2.35 fix PHP2Shell but not phpAfterShell), or your distribution's fixed package, then restart Apache and PHP-FPM. SOCRadar's advice is to verify the version the web server actually loads, not just php -v on the command line, which can differ. Include container images, Nextcloud and other appliances that ship their own PHP. - Treat anything exposed since 21 October as possibly compromised. CISA's KEV entry carries a forensic-triage flag for a reason. Patching does not remove an attacker who is already in. - Hunt for the persistence SOCRadar describes. Unfamiliar authorized_keys entries for the web user and root; cron jobs or /tmp/beacon.sh scripts that open outbound shells; PHP files you did not deploy in upload or document paths; an auto_prepend_file directive you did not set; a Dropbear SSH server bound to 127.0.0.1; web-server processes starting sh, curl or bash -i; and gitleaks, trufflehog or linPEAS running on a web server. Rebuild compromised hosts from clean images where you can. SOCRadar's post has the full indicator list to block and search. - Rotate every secret an exposed host could read. Cloud keys and role sessions, database credentials, .env values, OAuth and JWT signing keys, SSH keys and PHP session stores, then review cloud audit logs back to 21 October for enumeration of identity, storage, secrets, parameter and database services from unexpected addresses. Plan the order with the [Key Rotation Planner](/tools/key-rotation-planner) and work the incident with the [leaked credential response](/tools/leaked-credential-response) steps. - Close the cloud metadata path. Enforce IMDSv2 on EC2, narrow ECS task roles, and alert when a web process calls the metadata endpoint, as SOCRadar recommends. - Lock down Redis and other data stores. The largest completed theft in this campaign came from a Redis instance with no authentication. Require authentication, bind data stores to private networks and check nothing answers on the internet. - Limit what one web request can do. Cap multipart fields, parts and headers and rate-limit large POSTs at the proxy or web server, and turn off display_errors in production. These are PHP's and SOCRadar's partial measures, not a fix. - Prioritise with data and record it. Both CVEs are in KEV: check them in the [KEV checker](/tools/kev-checker/CVE-2026-109618), set deadlines with the [SSVC calculator](/tools/ssvc-calculator), and log hosts you cannot patch this week in your [risk register](/tools/risk-register) with an owner and a date. To see where AI-driven offence is heading, read our coverage of [ARTEX and agentic offensive frameworks](/blog/artex-agentic-offensive-frameworks-gtig) and the [agentic pentesting master guide](/blog/agentic-pentesting-systems-master-guide). For the PHP project's entry, with this campaign and the October fixes, see [PHP in the Breach Index](/breach-index/php). PHP2Shell is a vulnerability in PHP, not a breach of the PHP project. ## Frequently asked questions ### What is PHP2Shell? PHP2Shell is the nickname for two critical flaws in PHP itself, CVE-2026-109618 (remote code execution via out-of-memory corruption) and CVE-2026-109619 (remote code execution via stale global state). PHP fixed both on 21 October 2026 and rates each CVSS v4.0 9.2. PHP says the risk is highest under Apache with mod_php, where an unauthenticated request can trigger the attack before any application code runs. ### Which PHP versions fix PHP2Shell? PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, released on 21 October 2026. Every 8.2, 8.3, 8.4 and 8.5 release before those is affected, per the CVE records PHP published. PHP also provided minimal source backports for 7.4, 8.0 and 8.1, and Linux distributions shipped their own patched packages, so check your vendor. On 19 November 2026 PHP shipped 8.5.13, 8.4.28, 8.3.37 and 8.2.36, which fix further critical flaws including phpAfterShell (CVE-2026-126714), a session bug that its researchers say bypasses the PHP2Shell fix. Take those, the newest release on your branch. ### Is PHP2Shell being exploited? Yes. CISA added both CVEs to its Known Exploited Vulnerabilities catalog on 26 October 2026 with a 29 October due date for federal agencies, and its enrichment of the CVE records says exploitation is active. SOCRadar reported on 18 November that one operator exploited the flaws against exposed PHP/Apache servers from 21 October, the day public proofs of concept appeared. ### How did AI agents exploit PHP2Shell in the SOCRadar campaign? SOCRadar says the operator fed eligible targets into a self-built job queue that ran 39 parallel AI agent workers backed by two self-hosted DeepSeek V4.1 model endpoints. The workers checked targets, ran the exploit, created records in the operator's command-and-control server and ran data collection. SOCRadar calls the campaign human-directed and AI-augmented: it could not confirm how much was autonomous, and says no chain completed without manual oversight. ### How many servers were compromised by the PHP2Shell campaign? SOCRadar says the operator targeted 595 unique IPs in 57 countries and that 404 hosts are confirmed compromised. Interactive shells were documented on 217 hosts and root on only 12; most access ran as the unprivileged web-server user inside containers. As of 3 November about 290,000 files (9.5 GB) had been exfiltrated. ### Was any cryptocurrency stolen? No theft was confirmed, according to SOCRadar. The operator reached account-takeover and withdrawal capability on payment platforms, but a withdrawal of about 456,900 USDC stayed pending and an attack on a second platform failed because it could not reach the live system. SOCRadar says the completed impact on that track was about 3.3 million user and session records taken from an unauthenticated Redis instance. ### Who is behind the PHP2Shell campaign? Nobody has been named. SOCRadar assesses with moderate confidence that a financially motivated operator working like an initial-access broker ran it, and says its confidence in the operator's identity is low. Other firms have published separate reports on PHP2Shell activity; those are different observations and are not part of SOCRadar's figures. ### Is patching PHP enough after PHP2Shell? No, if the server was exposed before you patched. Updating closes the entry point, but SOCRadar says the operator left webshells, SSH keys, cron beacons and tunnels behind, and stolen cloud keys and application secrets stay valid until rotated. Hunt for persistence, rebuild compromised hosts where you can, rotate every secret the host could read and review cloud logs back to 21 October 2026. ### How can I tell whether my server was compromised? SOCRadar lists the signs: a /tmp/beacon.sh file or /tmp/.c marker files, a /var/www/html/.p/ directory, a Dropbear SSH listener on 127.0.0.1:2222, an auto_prepend_file setting you did not add, unfamiliar entries in authorized_keys, PHP files you did not deploy in upload or document paths, and web-server processes starting sh, curl or bash. Its post has the full indicator list. ### Was PHP or php.net hacked? No. Nothing reported shows a breach of php.net or the PHP project. PHP2Shell is a vulnerability in PHP, and the victims in SOCRadar's report are organisations running unpatched PHP servers. ## Sources - [php.net news archive 2026: security advisory and releases 8.5.12, 8.4.27, 8.3.36, 8.2.35 (21 October), informational advisory (23 October), release pre-announcements (20 October, 18 November)](https://www.php.net/archive/2026.php) - [PHP security advisory GHSA-w8hw-w9fx-j79w (CVE-2026-109618)](https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w) and [GHSA-6wmf-4rv5-p3rp (CVE-2026-109619)](https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp) - [CVE.org: CVE-2026-109618](https://www.cve.org/CVERecord?id=CVE-2026-109618) and [CVE-2026-109619](https://www.cve.org/CVERecord?id=CVE-2026-109619) (CNA PHP Group; affected ranges; CISA enrichment "Exploitation: active") - [NVD: CVE-2026-109618](https://nvd.nist.gov/vuln/detail/CVE-2026-109618) and [CVE-2026-109619](https://nvd.nist.gov/vuln/detail/CVE-2026-109619) (CVSS v4.0 9.2 from PHP) - [CISA Known Exploited Vulnerabilities catalog: both CVEs added 26 October 2026, due 29 October](https://www.cisa.gov/known-exploited-vulnerabilities-catalog?field_cve=CVE-2026-109618) - [PHP 8 ChangeLog (8.5.12, 8.5.13 and the other branches)](https://www.php.net/ChangeLog-8.php) - [PHP security advisory GHSA-fm2m-fjf5-v3cp (CVE-2026-126714, phpAfterShell), 19 November 2026](https://github.com/php/php-src/security/advisories/GHSA-fm2m-fjf5-v3cp) and [GHSA-g3vh-838c-65v7 (CVE-2026-129481)](https://github.com/php/php-src/security/advisories/GHSA-g3vh-838c-65v7) - [SOCRadar Threat Research Unit: PHP2Shell Exploited by AI Agents from Day One: Inside a Credential-Harvesting Campaign (18 November 2026), source of every campaign figure and the indicator list](https://socradar.io/blog/php2shell-ai-agents-credential-harvesting/)