What happened
SOCRadar’s Threat Research Unit says it identified attacker infrastructure associated with exploitation of PHP2Shell, tracked as CVE-2026-109618 and CVE-2026-109619. The campaign’s first exploit script was written on October 21, 2026, the day public proofs of concept appeared. The operator then adapted the exploit for default PHP/Apache configurations and individual PHP builds. SOCRadar says the operator worked from a pool of 68,198 PHP/Apache hosts identified through Shodan and attacked 595 unique IP addresses across 57 countries. Its report counts 404 confirmed compromises.
The campaign used a self-built job orchestrator called maq to queue work for patched Oh My Pi (omp) workers. SOCRadar reports 39 parallel workers and two SSH-forwarded, self-hosted vLLM endpoints running DeepSeek V4.1. The operator organized each target as a job, with a workflow spanning eligibility checks, exploitation, persistence, credential collection and upload. Its account describes the system as pursuing a goal of more than $50,000 USDT in payouts. That directive is evidence of an operator objective, not evidence that the target amount was collected.
SOCRadar says the operator’s collection pipeline took cloud API keys, IAM credentials, database backups, application secrets, OAuth tokens, repositories and session files. The report estimates the exfiltrated material at about 290,000 files and 9.5 GB. It also describes several mechanisms for maintaining access, including SSH keys, periodic reverse-shell beacons, token-authenticated webshells and encrypted Dropbear tunnels. These are the report’s account of what the operator deployed; the figures should not be read as a count of distinct people affected or as proof that every collected secret was used.
The financial-fraud track reached withdrawal and account-takeover capabilities on payment platforms, according to SOCRadar, but the report says it found no confirmed cryptocurrency theft. It cites 456,900 USDC in pending withdrawals and 7.3 BTC that remained encrypted. Separately, it reports that data taken from an unauthenticated Redis instance included about 3.3 million user records. SOCRadar also says an AI worker accidentally ran a collector against the operator’s own lab and uploaded a snapshot to the public loot service, exposing material about the operation.
What the archive adds
New Runtime’s earlier coverage documented PHP2Shell activity from two other angles: Delft honeypot data counted 72,343 requests from 659 IPs, while noting that requests do not equal successful compromises. SOCRadar’s account is a new fact on the same vulnerability story: it describes a specific operator’s workflow, reported compromises and data collection, rather than measuring honeypot traffic. New Runtime also recorded an expectation that exploitation would grow while internet-facing hosts remained unpatched; this campaign is consistent with that concern, but does not establish its future trajectory.
Trend signal
This report is consistent with the concern that PHP2Shell exploitation will continue while unpatched internet-facing PHP hosts remain, as described in New Runtime’s earlier assessment. It adds a reported example of automation being applied across target selection, job execution and follow-on collection. One campaign cannot establish how common AI-agent orchestration is among attackers, or whether it materially increased their success rate. This supports the vulnerability-exploitation concern, not a broad claim that agent-driven attacks are now typical.
Why now
SOCRadar dates the report to November 18, 2026, and places the operator’s first exploit script on October 21, the day public proofs of concept were released. Its account therefore describes activity beginning at disclosure and continuing into the following weeks, including a second exposed directory discovered on November 8. The timing makes the report relevant to teams still assessing exposure and reviewing whether vulnerable PHP services were patched promptly.
Why it matters
For teams building with AI, the operational lesson is not that an agent independently devised this entire intrusion. SOCRadar describes an operator-built system that used agents as workers inside a larger pipeline: targets were sourced and screened, work was queued, exploit attempts ran, and harvested material was collected. Automation can multiply the reach of a workflow even when humans or prebuilt scripts define its objectives and components.
The reported collection also shows why server compromise should not be treated as a contained application incident. Cloud credentials, repositories, database backups and session files can create routes to other systems and users. Teams should therefore connect vulnerability response with credential rotation, cloud audit review and checks for unauthorized persistence. The report does not show that every listed follow-on consequence occurred, but its account makes the potential blast radius concrete: stolen access material is itself a high-value outcome, even where cryptocurrency theft is unconfirmed.
Limits and uncertainty
These figures are SOCRadar’s findings from its analysis of exposed operator infrastructure and associated records; they are not an independently verified census of all PHP2Shell activity. “Confirmed compromised” is the report’s classification, and the available account does not establish how many distinct organizations or people were affected. The 290,000 files and 9.5 GB describe collected material, not necessarily unique data or data that was all successfully exploited.
The report documents an agent-orchestrated workflow but does not establish how much the agents contributed compared with the operator’s scripts, configuration and direction. Nor does the stated payout goal show that the goal was met. SOCRadar reports pending withdrawals and encrypted bitcoin rather than confirmed cryptocurrency theft. The extent of downstream use of stolen credentials and records remains uncertain.
Reader takeaway
Treat an exposed vulnerable service as a possible starting point for credential and data theft, not merely as a code-execution problem. Patch affected PHP deployments, then review access and persistence: rotate secrets that may have been exposed, inspect cloud activity and check for unauthorized keys, scheduled tasks and webshells. SOCRadar’s report is one campaign account, not a measure of prevalence—but it illustrates why remediation and post-compromise review belong together.