CourseLarge Language Models: The Complete Course · Module 11: Safety, Security, and Alignment · part 82 of 103
Part 82 · Module 11: Safety, Security, and Alignment

Incident: A chatbot reply exposed support agents' sessions

7 min read·13 Oct 2026
Module 11 · Incident file · About 25 minutes

What happened

LENOVO'S LENA, AS CYBERNEWS REPORTED IT one prompt 400 characters chatbot GPT-4, complies chat page renders markup outside server gets the cookie support agent after handover with that cookie: the support system, as the agent
The attack (red) went through the model (amber) and a page that rendered its markup (grey), reaching the human agent who took over the chat.

On 18 August 2025 the Cybernews Research team published "Critical flaw plagues Lenovo AI chatbot: attackers can run malicious code and steal cookies" (Cybernews, 18 August 2025). The chatbot was Lenovo's customer support assistant, Lena (no relation to Quillpath's agent Lena), which the article describes as "powered by OpenAI's GPT-4."

The researchers wrote that "All it took to exfiltrate active session cookies from Lenovo was a single 400-character-long prompt". The chatbot was made to write its answer as web markup, and that markup made the reader's browser send a request to an outside server, carrying the session cookie. The article stresses the handover: "Once you're transferred to a real agent, you're getting their session cookies as well". With a support agent's cookie, it says, someone could log in to the customer support system as that agent, and "potentially access active chats with other users and possibly past conversations and data".

DISCLOSURE TIMELINE, 2025 (CYBERNEWS) 22 Jul reported 6 Aug acknowledged before 18 Aug mitigated 18 Aug published no exploitation or cost was reported
Cybernews reported the flaw the day it found it, and published after Lenovo had mitigated it; no harm or cost was reported.

Cybernews gives the timeline: found and reported to Lenovo on 22 July 2025, acknowledged on 6 August, mitigated before 18 August, published on 18 August. CSO Online (Gyana Swain, 20 August 2025) quotes Lenovo's spokesman: Lenovo "promptly assessed the risk and implemented corrective actions." Neither source reports that anyone exploited the flaw or what it cost; the harm described is what an attacker could have done.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside