Loading story…
An Information Security Media Group interview with security researcher Candid Wüest (Xorlab) discussed how generative AI systems (e.g., ChatGPT, Gemini, Claude) can be used to produce malware and ransomware code, but typically include guardrails intended to block overtly malicious requests. Wüest said attackers can still obtain harmful output through prompt “jailbreak” techniques, where the model is asked indirectly or with obfuscation rather than with explicit malware-writing instructions.
The interview highlighted that some attacker-oriented or less-restricted models (e.g., WormGPT) reduce the need for rephrasing, and described a current evasion pattern of requesting malicious content in poetic/rhyming formats (e.g., limericks/sonnets) to coax responses that would otherwise be refused. The content is presented as expert commentary on attacker tradecraft and AI safety limitations rather than reporting a specific breach, vulnerability disclosure, or active campaign.
Track how attackers are adapting to this technology.
4 events from the most recent confirmed update back to the earliest known activity.
In interviews published on February 10, 2026, Wüest said generative AI can help attackers create malware and bypass some guardrails through jailbreaks or stepwise prompting, but does not reliably produce novel, undetectable ransomware. He emphasized that behavior-based defenses still detect actions such as file encryption and warned about enterprise AI copilots creating data leakage and destructive automation risks.
Vendor reporting cited by Wüest said a Chinese nation-state actor used Claude to automate large portions of an intrusion chain. He noted the details were limited and that hallucinations reportedly caused failures and false claims of access.
Following reporting on LameHug in 2025, Candid Wüest said the malware was likely associated with the Russian state-linked threat group APT28. This represented an attribution assessment tied to the AI-assisted malware example.
In 2025, Ukrainian CERT reported LameHug, an early malware example that embedded an LLM prompt and used Qwen 2.5 to generate system-discovery commands. Xorlab's Candid Wüest described the implementation as having little practical variability because of a low temperature setting.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Check Point Research reported that cybercriminals have started experimenting with **ChatGPT** to support malicious activity, using the tool to generate convincing phishing lures, draft fraudulent business communications, and assist with coding tasks tied to malware development. The report highlighted examples from underground forums showing threat actors discussing how large language models could lower the barrier for less-skilled operators by helping produce polished English text and functional code snippets. Researchers said the observed use was still early-stage, but the activity showed how generative AI could accelerate common cybercrime workflows, including social engineering, scripting, and evasion-oriented development. The findings underscored a growing risk for defenders: AI tools are being adapted not only for productivity and research, but also to improve the speed, scale, and professionalism of criminal campaigns.
Jun 29, 2026
Security researchers and vendors are warning that **prompt injection and jailbreak techniques** remain a leading risk for enterprise deployments of large language models (LLMs), enabling attackers to override system instructions, bypass safety controls, and potentially drive **data exposure** outcomes. Resecurity reports assisting a Fortune 100 organization where AI-powered banking and HR applications were targeted with prompt-injection attempts, emphasizing that these attacks exploit model behavior rather than traditional software flaws and can be used in scenarios such as extracting sensitive configuration data (for example, attempts to elicit content resembling `/etc/passwd`). Resecurity also cites OWASP’s 2025 Top 10 for LLM Applications, where prompt injection is ranked as the top issue, and frames continuous security testing (e.g., VAPT) as a key control for enterprise AI systems. Separate research highlighted by Kaspersky describes a **“poetry” jailbreak** technique in which prompts framed as rhyming verse increased the likelihood that chatbots would produce disallowed or unsafe responses; the study tested this approach across 25 models from multiple vendors (including Anthropic, OpenAI, Google, Meta, DeepSeek, and xAI). In contrast, OpenAI’s planned upgrade to *ChatGPT Temporary Chat* is primarily a product/privacy change—adding optional personalization while keeping temporary chats out of history and model training (with possible retention for up to 30 days)—and does not describe a specific security incident or vulnerability disclosure tied to prompt injection or jailbreak research.
Jun 30, 2026
New research highlights how **large language models (LLMs)** can be operationalized for offensive use, including generating malicious JavaScript and exploit code with limited human involvement. Unit 42 described an *AI-augmented runtime assembly* technique in which a seemingly benign webpage makes client-side API calls to trusted LLM services to obtain code fragments that are then assembled and executed in the victim’s browser, producing a personalized phishing experience. The approach is designed to be evasive by delivering content from trusted LLM domains, producing polymorphic code per visit, and deferring malicious behavior until runtime—reducing the effectiveness of static and network-only detections. Separately, an experiment reported by CybersecurityNews described testing GPT-5.2- and Opus 4.5-based systems against a **zero-day** in the *QuickJS* JavaScript interpreter, resulting in **40+ distinct exploits** across multiple configurations and protection scenarios. The report claims GPT-5.2 solved all presented challenges and that many exploit-generation runs completed in under an hour at relatively modest token costs, suggesting exploit development could increasingly scale with compute and budget rather than scarce expert labor. Together, the reports reinforce that LLMs can be used both for **client-side phishing payload generation** and for **automated vulnerability exploitation**, increasing the speed and variability of attacks defenders may face.
Jun 29, 2026