Oracle EBS CVE-2025-61882: Cl0p Extortion Campaign

What does CVE-2025-61882 allow?

NVD describes an "easily exploitable" vulnerability in Oracle Concurrent Processing (BI Publisher integration): an unauthenticated attacker with HTTP network access can take over the component. Versions 12.2.3 to 12.2.14 are affected. BleepingComputer notes that the October 2023 quarterly patch must be installed before applying the new fixes (Oracle note 3106344.1).

According to Google Threat Intelligence (GTIG) and Mandiant, a leaked exploit chain combines SSRF, CRLF injection, authentication bypass and XSL template injection. The observed mechanism: a malicious XSL template is stored in the database (table XDO_TEMPLATES_B) through the XDO template manager, then triggered through the TemplatePreviewPG page. GTIG says it remains hard to tell which specific flaws map to this CVE.

What is the timeline?

• July 2025: suspicious activity on /OA_HTML/configurator/UiServlet; Oracle fixes flaws in its July update.
• August 9, 2025: first confirmed exploitation of what may be CVE-2025-61882, via SyncServlet.
• September 29, 2025: GTIG and Mandiant begin tracking an extortion email campaign sent from hundreds, if not thousands, of compromised third-party accounts.
• October 2, 2025: Oracle warns that flaws fixed in July may have been exploited.
• October 4, 2025: emergency patch for CVE-2025-61882; October 6: KEV listing.
• October 11, 2025: patch for CVE-2025-61884.

Who is behind the campaign?

GTIG does not formally attribute the activity to a tracked group. The Cl0p brand and the contact addresses listed on Cl0p's leak site appear in the emails, but the brand has been used by several actors. Mandiant later said Cl0p exploited the flaws, and Cl0p told BleepingComputer it used an Oracle zero-day. The victims cited in the trade press (Harvard, Envoy Air, the Washington Post, Logitech) come from public confirmations; other names seen in secondary sources, including French manufacturers, rest on no primary confirmation and are not repeated here.

Who is affected?

Organizations running Oracle E-Business Suite 12.2.3 to 12.2.14, especially those whose instance is reachable from the internet. An ERP centralizes payroll, supplier contracts, consolidated finance and personal data: a pre-authentication flaw gives direct access to the heart of the company. An instance exposed between July and October 2025 should be treated as potentially compromised.

How do you detect exploitation?

1. Query the database: SELECT * FROM XDO_TEMPLATES_B ORDER BY CREATION_DATE DESC; and the XDO_LOBS table, looking for templates whose code starts with TMP or DEF.
2. Search logs for requests to TemplatePreviewPG, UiServlet and SyncServlet, and the TemplateCode=TMP pattern.
3. Examine processes: bash -i launched by the applmgr Java process, and in-memory implants (memory analysis of EBS Java processes).
4. Compare outbound traffic with the indicators published by Oracle and Google, and block non-essential internet access from EBS servers.

How do you fix it?

Apply the October 2025 emergency patches (and the one for CVE-2025-61884), after the October 2023 quarterly patch if needed. Take the ERP off the internet behind authenticated access and strict network filtering, restrict outbound traffic, rotate the instance's secrets and watch for extortion emails addressed to executives: the stolen data, dated mid-August 2025 in some cases, serves as proof.

What changes when an AI agent has access to an ERP?

Finance and procurement departments deploy agents that read the ERP: invoice matching, supplier reminders, journal entry preparation, natural-language questions about payroll. These agents receive a broad integration account. A compromised ERP gives the attacker that account and the content the agent reads; the attacker can also edit free-text fields (invoice labels, order comments) to plant an instruction the agent will carry out (OWASP LLM01).

A structural parallel deserves attention: the attack stores an executable template in the database, then triggers it later. Agent systems do the same when prompt templates or tool instructions are stored in a database or shared repository: whoever can write them can change what the agent does, hence the importance of controlling write rights on those items (LLM01, LLM03). Other effects: the concentration of sensitive data reachable by the agent (LLM02), the speed of automated extraction and phishing: extortion emails sent from real compromised accounts look like internal messages, which mail agents treat as reliable. Useful controls: one integration account per agent with limited read-only rights, ERP writes only with human approval, restricted write rights on templates and instructions, extraction ceilings and logging. See AI agent security and the neighboring article on Windchill and FlexPLM.

Which agent tests apply to Oracle EBS CVE-2025-61882?

An agent red team plants a booby-trapped free-text field in the data the agent reads, and checks what it executes, what it reveals and what volume it can extract. On the offer side, the Agent Red Team Express simulates the data loss of an agent connected to Oracle EBS (5 days, 1,900 euros). The method details the approach.

What is the lesson?

An exposed ERP with a pre-authentication flaw produces mass leaks within weeks. Take it off the internet, patch on publication, look for traces over the period preceding the fix, and treat the agents plugged into it as privileged accounts.