exploit_agent: Automating Everything from Patch Diff Analysis to Exploit Generation and Execution Without Human-in-the-Loop
Basic Information
Article: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold
Published by: Google Threat Intelligence Group (GTIG) / Mandiant
Publication Date: 2026-11-16
Original Source: Google Cloud Threat Intelligence
Related Source: BleepingComputer
Related Malware: exploit_agent
Related Threat Groups: Unspecified, PRC-nexus group (observed developing a separate custom framework)
Related CVEs: No specific CVEs have been identified for exploit_agent or PentAGI activity.
Related Products: Spectra (Ultimate Addons for Gutenberg), PaperCut, PentAGI, CAIRN, CyberStrikeAI, HexStrike AI
Importance (author-assessed response priority): High
In Brief
GTIG identified a multi-agent framework named exploit_agent operating on a public staging server.
The framework continuously compares web application code before and after patches, analyzes potential vulnerabilities, creates exploits, and automatically executes them against targets without human-in-the-loop intervention.
The server also contained data gathered from victims, along with scripts for internal reconnaissance, brute force, and web shell deployment.
Attack Flow
Automated Vulnerability Weaponization by exploit_agent
- The framework continuously collects patch announcements and source code changes for target web applications before and after updates.
- It uses multiple frontier and local LLMs to analyze diffs, extract potential vulnerabilities, and identify exploitable paths.
- It automatically generates exploit code and runs it against target servers without human-in-the-loop intervention.
- It aggregates retrieved data on a staging server.
GTIG identified victim telemetry showing interest in French media, entertainment, and education sites.
Confirmed Related Capabilities: GTIG identified scripts on the staging server capable of supporting internal reconnaissance, brute force, web shell deployment, and the removal of other attackers' web shells.
Execution methods and sequences within individual victim environments have not been publicly disclosed.
Reconnaissance Against Organizations in Asia Using PentAGI-Based Orchestration
Attackers built a custom attack-orchestration platform based on open-source PentAGI.
The platform executed systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia.
Public information confirms exploration activity from late August to early October 2026, but does not indicate the number of successful compromises.
Attacker Location and Execution Venue
External attackers connecting from staging and orchestration infrastructure on the internet to public web applications, servers, and APIs.
The framework can combine external LLM APIs with local models to automate patch diff analysis, exploit generation, and the ingestion of execution results.
GTIG identified scripts supporting internal reconnaissance, brute force, and web shell deployment and removal on the staging server.
The execution locations, scope, and sequence in individual victim environments have not been publicly disclosed.
Visibility for Victims and Administrators
Users: Public web services may continue to function normally while data exfiltration or web shell deployment progresses in the background.
Administrators: Inference: In environments with appropriate logging enabled, activity may appear as unnatural diff-dependent requests immediately following patch disclosure, repeated scans and payload modifications from the same source, shell execution or brute-force tool runs originating from web processes, the creation or removal of web shells, and data transmission to external staging destinations.
Reports do not indicate that all of these indicators were observed in every victim environment.
Success and Failure Conditions
Success Conditions
Attackers must be able to access comparable pre- and post-patch code, release notes, or PoCs.
Internet-facing applications must remain unpatched and accept payload execution and result feedback from the framework.
Web process privileges and network reachability must allow internal reconnaissance, credential access, web shell deployment, and data transmission.
Failure Conditions and Mitigations
Promptly apply vendor patches and limit the exposure of release processes, including preview or pre-release branches and early diffs.
Run web applications with the principle of least privilege, egress restrictions, read-only filesystems, and network segmentation to reduce the blast radius after an initial exploit.
Use WAFs, rate limiting, and EDR to block repeated payloads and abnormal child processes spawned by web processes, and channel blocked attempts into follow-up investigations.
What Happens Upon Success
Reduced time from vulnerability weaponization to exploit execution against public web applications.
Internal reconnaissance, brute force, web shell deployment, and removal of competing attackers' web shells on web servers.
Acquisition of victim data and aggregation onto external staging servers.
Compressed defense windows before and after patch application due to parallel scanning across numerous targets.
Observable Logs
Inference: The following items are candidates for investigation. An attacker's use of external LLMs or retrieval of public source code may not leave traces in victim-side logs. Access to source repositories and CI artifacts should be verified using audit logs restricted to organization-managed assets.
Email: No specific notes.
Proxy / SWG / DNS: Check for repeated requests to the same app or plugin endpoint immediately after disclosure, unknown orchestration hosts, external LLM or routing services, and egress traffic to staging destinations. LLM API traffic alone does not confirm an attack.
Endpoint / EDR: Check for shells spawned by web or PHP processes, scanners, brute-force tools, new files created in the web root, deletion of existing web shells, and access to source or patch artifacts.
Identity / IdP: Check for brute-force attacks against web applications or CMS accounts, subsequent session and administrator creation, and the use of service credentials.
SaaS / Cloud: Check for retrieval of source repository contents before and after patches, CI artifact access, secret reads on cloud-hosted web workloads, and suspicious egress traffic.
Network: Check for repeated requests to the same target with varying payloads, and short-duration sequences of scanning, exploitation, internal communication, and data egress.
Determining Attack Success
Confirmed in Public Information
Attack Attempts Observed (Success Unconfirmed): GTIG reported that a custom PentAGI platform performed systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia between late August and early October 2026. Public information does not indicate the number of successful compromises resulting from this activity.
Data Theft or Session Compromise Confirmed: GTIG confirmed data retrieved from targets on an exploit_agent staging server. However, the number of victim organizations, independent verification for each organization, and the scope of data retrieval have not been publicly disclosed.
Criteria for Internal Assessment
Do not treat patch diff retrieval or scanning alone as confirmation of success; verify exploit responses, web process child execution, file creation, credential access, internal communication, and data egress step by step.
Correlate repeated payloads, execution feedback, and short-duration attack chains rather than relying solely on framework names or LLM APIs.
Investigation Playbook
Investigation Starting Point: Begin with repeated scans immediately after patch disclosure, shells spawned by web processes, web shell creation or deletion, brute-force traffic, and egress to unknown staging infrastructure.
Initial Verification: Establish a chronological timeline of target versions, patch application times, initial requests, responses, sources, WAF results, and process or file modifications.
Endpoint and Server Investigation: Preserve web roots, temporary directories, process trees, shell history, credential files, web shells, and deleted artifacts.
Authentication and Cloud Investigation: Verify CMS or application logins, new admin accounts, tokens, secret stores, source repositories, and CI artifact access.
Follow-up Activities: Track internal scanning, brute-force attempts, credential reuse, web shells, data collection, and staging or egress traffic.
Containment: Restrict network access to affected applications, apply patches or rebuild, rotate credentials, remove web shells, and block egress traffic.
Classification Categories: Differentiate code diff retrieval, scanning, exploit transmission, command execution, web shells, credential access, data acquisition, and egress.
Defense and Detection Ideas
Single Events: Treat shells spawned by web processes, executable files created in web roots, deletion of existing web shells, and unknown payloads arriving shortly after patch disclosure as high priority.
Chronological Correlation: Correlate release or patch disclosure, diff-dependent scanning, payload repetition, command execution, internal reconnaissance, and staging or egress sequences.
Threat Hunting: Search for unpatched internet-facing plugins, CMS instances, web process child processes, recent web shells, brute-force tools, and sources sending uniform requests to numerous hosts in a short period.
Log Limitations: Because the framework operates on attacker infrastructure, victim environments may not retain traces of exploit_agent or model names. The absence of tool names does not rule out automated attacks.
Priority Mitigations: Prioritize internet-facing asset inventories, rapid patching, WAF virtual patching, least-privilege execution and egress restrictions for web workloads, and the centralization of process, file, and application logs.
Facts / Inference / Hypothesis
Facts
GTIG identified an exploit_agent staging server that integrates multiple LLMs to automate patch diff analysis, vulnerability candidate extraction, exploit generation, and automatic execution against targets.
GTIG reported that this pipeline fully automated vulnerability discovery and exploitation without human-in-the-loop intervention.
The server contained data retrieved from victims along with scripts supporting internal reconnaissance, brute force, web shell deployment, and the removal of competing attackers' web shells.
Victim telemetry indicated interest in sites associated with French media, entertainment, and education.
GTIG independently observed the use of CAIRN and PentAGI in the wild, as well as trials of CyberStrikeAI and HexStrike AI.
A custom PentAGI platform conducted reconnaissance and vulnerability scanning against a wide range of organizations in Asia between late August and early October 2026.
GTIG described some activity as PRC-nexus, which does not prove Chinese government direction or involvement.
GTIG referenced PaperCut CVE-2026-81578 / 82078 and PHP CVE-2026-109618 / 109619 as separate examples. The report does not state that the exploit_agent or PentAGI activity in this article exploited these specific vulnerabilities.
Inference
The most critical defense window begins not only after patch publication, but as soon as pre-release code or early patch diffs become available.
Sequences involving scans triggered by patch diffs, payload repetition, web process anomalies, and egress serve as more consistent detection axes than framework-specific IOCs.
Hypothesis
No additional hypotheses. Do not assert victim counts, attribution, or exploit success rates beyond publicly available information.
MITRE ATT&CK Mapping
T1595Active Scanning - high - Systematic reconnaissance and vulnerability scanning were observed originating from the PentAGI-based platform.T1190Exploit Public-Facing Application - high -exploit_agentgenerates and executes exploits using patch diffs of web applications.T1110Brute Force - medium - Post-compromise scripts supporting brute force were identified on the staging server, though execution scope in individual victim environments remains undisclosed.T1505.003Server Software Component: Web Shell - medium - Web shell deployment and removal scripts were identified, though the scope of successful deployment across individual targets remains undisclosed.
Open Questions and Further Investigation
The operators of exploit_agent, the number of victim organizations, all exploited vulnerabilities, exploit success rates, and the scope of data acquisition.
The extent to which web shell deployment, brute force, and internal reconnaissance scripts were actually executed in individual victim environments.
Operational and infrastructural relationships between the PRC-nexus custom framework and PentAGI / exploit_agent.
Complete mapping relationships for activities utilizing CAIRN, PentAGI, CyberStrikeAI, and HexStrike AI.
Impact on SOCs and Organizations
Because PentAGI activity targeting Asia has been confirmed, organizations should treat internet-facing CMS platforms, plugins, and business applications as priority assets.
Operating under the assumption that the window between patch diff publication and exploitation is compressed, organizations should combine rapid patching, virtual patching, web process monitoring, outbound traffic restrictions, and source or CI artifact access auditing.
Audience-Specific Summary
For SOCs: Correlate scans immediately following patch disclosure, payload repetition, web process child activity, web shells, credential access, and staging or egress traffic on an asset-by-asset chronological basis.
For Administrators: Promptly update public web applications and reduce the blast radius after initial exploits using WAFs, the principle of least privilege, read-only filesystems, egress restrictions, and secret isolation.
For Users: This incident prima
Comments
No comments yet. Start the discussion.