Evidence guide: input, decoy response, and runtime proof

Deception Check · October 23, 2026 · Static reconstruction, not exploit replay

Use this guide to turn the observed requests into an investigation: find the destination and source, identify the serving process, and correlate its behavior. The 26-record selection contains 23 PHP2Shell tool-labelled probes, one sender-tagged downloader input, and two separate Extensive VC inclusion attempts. Read the research article.

Start with these three checks

  1. Search destination 217[.]60[.]102[.]5 in egress records and /sh in visible HTTP URLs. Correlate with endpoint process and command-line context. Search libredtail-http in incoming HTTP logs and 77[.]90[.]185[.]20 as a source in both web and SSH records.
  2. Identify the actual service model: Apache mod_php, FPM, CGI, or another SAPI. Check direct shell/downloader children, worker restarts, crashes, and memory errors. Preserve the process identity, request interval, deployed build, and clock basis.
  3. Compare with normal application behavior. A worker error or outbound event is a lead; correlate response bytes, file changes, and process ancestry before concluding execution. The on-page detection appendix provides separate request, runtime, destination, and error-log hunts.

The evidence ladder

Swipe tables to see all columns.

Layer Evidence retained here Claim it permits
Network receipt Timestamp, source, method, path, and selected body fields A request reached the decoy
Source join Request source, or same-sensor, same-session connection The recorded address submitted that request
Capture metadata Filename, byte count, hash, and quarantine tag The handler recorded a capture with those attributes
Static file bytes Six retained copies, two distinct verified hashes Exact repetitive filler content in a/b, not PHP execution
Decoy response Status, length, and generation flag The decoy recorded that response metadata
PHP runtime Not established for these events No execution or compromise conclusion

An event named malware_capture is a logging label, not a completed malware analysis. llm_generated:true identifies response provenance, not attacker input. llm_generated:false is also not proof that a PHP engine executed anything.

Sequence A: labelled multipart probe

  1. A same-session connection record identifies the source.
  2. Capture metadata records uploads named a and b.
  3. The request record contains a capture placeholder and a PHP2Shell user agent.
  4. The response records status 200 and metadata, not a PHP process trace.

The earliest request timestamp is October 21 at 08:42:42.137613 UTC. Capture events can precede the request record while the handler processes the upload; this log order is not a sequence of proven exploit stages. The preserved prefix from later web requests is insufficient to validate a working OOM or stale-state chain.

Byte-level a/b inspection

Three retained copies per artifact reproduced the complete file contents and recorded hashes. The following offsets describe the uploaded files, not the entire HTTP request.

Swipe tables to see all columns.

File Zero-based offsets Exact bytes
a 0 to 1 0d 0a (CRLF)
a 2 to 5,185 5,184 × 41 (A)
a 5,186 to 10,369 5,184 × 42 (B)
b 0 to 63 64 × 43 (C)

All six copies match their expected SHA-256 and MD5, including the hashes in earlier capture events. Their complete contents are repetitive ASCII filler. This establishes the file bytes, not the surrounding multipart body. Exact analysis and hashes.

VulnCheck interprets the matching a/b pattern as an attempt to exhaust memory, insufficient in one request in its observations. Treat the repetitive bytes as padding and inspect multipart shape, size limits, allocation errors, and worker behavior. A capture handler's text/script label does not change the bytes. VulnCheck.

Sequence B: separate Extensive VC inclusion attempts

Two separate targets received the same pattern from 45.115.26.26 on October 22:

Swipe tables to see all columns.

UTC interval Incoming behavior Recorded response
21:53:15 to 21:53:24 File-read probes, upload at 21:53:19.178708, then three inclusion requests File reads/upload 200; inclusion requests 404; generated responses
21:54:13 to 21:54:18 File-read probes, upload at 21:54:17.455252, then three inclusion requests File reads/upload 200; inclusion requests 404; generated responses

The 297-character upload body includes PHP in a filename, a session field, and an upload-progress field. The filename contains the marker EXTVC_OK and an attempted system() call reading $_GET['c']. The file content is Z.

Subsequent form bodies use the action extensive_vc_init_shortcode_pagination and options[template] to reference /tmp/sess_[redacted], /var/lib/php/sessions/sess_[redacted], and /var/lib/php/session/sess_[redacted]. The submitted c=id is POST form data; the logged query string is empty. Do not turn that into proof of an executed id command. A 200 response to a file-read probe is likewise not evidence that actual host files were disclosed by a generated decoy.

The action plus options[template] matches the Extensive VC Addons pattern documented for CVE-2023-0159 in Exploit-DB 52085. CVE-2025-14475 concerns a different shortcode_name parameter in the same action; it is not the appropriate mapping for these records. Keep this plugin technique separate from PHP2Shell. The two uploads remain in the 26-record selection; neighboring inclusion requests are contextual records.

Sequence C: submitted downloader command

At 21:55:25.787307 on October 22, 77.90.185.20 submits a multipart body containing PHP in PHP_SESSION_UPLOAD_PROGRESS. Safe semantic reconstruction:

INCOMING PHPPrint a start marker
REQUESTED ACTIONFetch an HTTPS script with wget or curl
REQUESTED ACTIONPipe bytes to sh with a php2shell argument
INCOMING PHPPrint an end marker

The destination is hxxps://217[.]60[.]102[.]5/sh. The POST records 404 and the adjacent session-path GET records 200, both with zero-length generated responses. The sequence is an upload-progress/session-file inclusion attempt; php2shell is the submitted shell argument, not a validated exploit mechanism. Inspect destination events and the process that initiated them in real endpoint/network telemetry.

Selection and counting method

Selection uses PHP2Shell identifiers or upload-progress markers in incoming request fields, excluding generated responses. Twenty-eight matching rows reduce to 26 unique original request lines after exact-hash deduplication. Eleven contain a source address; 15 use a connection record with the same sensor identity and session. No timing-only or cross-sensor joins are used.

Classify after selecting: 23 tool-labelled probes/reconnaissance requests from three addresses; one php2shell-tagged downloader input from a fourth; and two separate Extensive VC uploads from a fifth. Three labelled requests have no retained bodies and stay separate from the multipart category. The public event index preserves stable IDs and original-line hashes so readers can trace the counts; archive custody and review records are retained privately.

Shared delivery campaign: counting requests and sessions separately

Four input routes reference the same defanged delivery endpoint. This is a semantic map of submitted inputs, not a diagram of executed processes:

HTTP TRAVERSAL104 distinct requests / 47 sources
SSH21 delivery sessions / 2 sources
TELNET16 delivery sessions / 15 sources
PHP UPLOAD PROGRESS1 request / 1 source

Each route requests hxxps://217[.]60[.]102[.]5/sh, with a corresponding apache, ssh, telnet, or php2shell argument. These labels describe incoming text; script contents and argument handling are unverified. The same source, 77.90.185.20, submits SSH delivery on October 21 at 21:25:51.087219 UTC and PHP delivery on October 22 at 21:55:25.787307 UTC.

For this separate campaign comparison, we scan incoming HTTP fields and command fields for the literal endpoint, excluding response text, summary/index recounts, and 24 synthetic download-URL enrichment records. Deduplication uses sensor identity plus exact raw JSON-line SHA-256 excluding line endings, preserving timestamps. That differs deliberately from the PHP2Shell public index's original-line hash including LF; the two hash labels are not interchangeable.

After deduplication and inclusion of recovered records, the result is 104 HTTP requests from 47 sources, 21 SSH sessions from two sources, and 16 telnet sessions from 15 sources. The SSH sessions contain 63 host-bearing command fragments, not 63 wire commands or attacks. Keep request and session counts separate. Machine-readable counts, windows, and reconciliation.

libredtail-http appears in all 104 HTTP traversal requests, not in the php2shell-tagged downloader request. Published SANS observations and Akamai's sample-based research support a RedTail-consistent assessment of delivery style. We did not acquire this endpoint's script or an installed binary, so family identification remains unconfirmed. The endpoint's observed role is delivery, not command-and-control. SANS, Akamai.

Indicator roles and safe use

The JSON indicator list distinguishes five observed source addresses, one submitted downloader destination, tool labels, contextual markers, request paths, and two byte-verified filler-artifact hashes. The plain-text version supports manual review. Network values are defanged for handling; normalization instructions are included for authorized log hunts.

Swipe tables to see all columns.

Signal Main limitation
Source IP Reassignment, proxies, research activity, and shared hosting; no actor identity
User agent Caller-controlled and easy to copy
Upload-progress field Can be legitimate; not a CVE fingerprint by itself
Boundary or output marker Tool/payload correlation, not execution proof
a/b artifact hash Exact filler bytes verified; correlation, not malware classification or exploit-success proof
PHP inside a filename or progress field Stronger evidence of code-injection intent; still not execution

Use the research detection appendix for source rules, platform hunts, telemetry requirements, and validation status. Combine broad strings such as upload-progress field names, filenames, or shell-tool names with suspicious structure and adjacent activity. Test the logic against your own telemetry schema and benign uploads before operational use.

What would close the execution question?

Obtain response bytes with provenance, a retrievable capture reference, and runtime telemetry from the actual affected service. Correlate PHP worker identity across requests, crashes or error handling, child process ancestry, file modifications, and egress. Preserve build/version evidence and clock quality. A shell response marker alone can be fabricated by a decoy and is not enough.

The two PHP advisories explain different failure mechanisms. OOM corruption can leave incompletely initialized state for cleanup. Stale global state requires request-owned state surviving into a subsequent request handled by the same worker. Similar-looking HTTP input does not establish either condition.