Loading story…
New research reports that LLM agents can deanonymize pseudonymous users by extracting biographical and behavioral clues from a small number of “anonymous” posts and then using web search and reasoning to identify real-world identities. The technique was evaluated across multiple sources of unstructured text (including Hacker News, Reddit, LinkedIn, and anonymized interview transcripts) and is described as achieving high precision while scaling to tens of thousands of candidates, making a task that previously required significant human effort increasingly practical.
Experimental results described in coverage indicate LLM-based approaches can outperform classical deanonymization baselines (including methods similar to the Netflix Prize-style linkage attack), with precision degrading more slowly as the attacker makes more guesses and with agentic steps (e.g., “Reason” and “Calibrate”) improving recall at high-precision thresholds. Proposed mitigations include rate limiting and anti-scraping controls on user-data APIs, restricting bulk exports, and LLM-provider monitoring/guardrails to detect and refuse deanonymization misuse; the research warns the capability could be abused by governments to unmask critics, by advertisers for hyper-targeted profiling, or by criminals to build target dossiers at scale.
Track how attackers are adapting to this technology.
3 events from the most recent confirmed update back to the earliest known activity.
Alongside the findings, the authors proposed defenses including rate limiting, anti-scraping detection, restricting bulk data exports, and LLM-provider guardrails to refuse deanonymization requests. They warned that improving LLM-based re-identification could aid government surveillance, hyper-targeted advertising, and personalized social engineering.
The research reported successful deanonymization across platforms including Hacker News, Reddit, LinkedIn, anonymized interview transcripts, and a Netflix-derived evaluation setup. In one reported test, the system achieved up to 67% accuracy at 90% precision on nearly 1,000 LinkedIn-to-Hacker News matches, and another dataset showed correct re-identification of 9 out of 125 candidates.
Researchers from ETH Zurich created an automated framework that uses large language models to infer identity-relevant details from pseudonymous online posts, search the web for matching profiles, and rank likely identities with confidence scoring. The work was designed to show that LLMs can automate deanonymization that previously required substantial manual effort.
Follow how adversaries are adapting to this technology, and where it touches your stack today.
4 references tracked. Mallory keeps watching after this page renders.
cyberscoop.com
Open sourcetechxplore.com
Open sourcearstechnica.com
Open sourceschneier.com
Open sourceMap 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.
Academic researchers demonstrated that **large language models (LLMs)** can *automate deanonymization* by building profiles from a person’s historical online comments and other unstructured “digital exhaust,” then linking those profiles across platforms even when different pseudonyms are used. In limited testing, the approach reportedly achieved **99% precision** when linking **Hacker News** accounts to **LinkedIn** profiles by correlating signals such as vocabulary, location hints, hobbies, and other biographical clues, undermining the long-held assumption that pseudonymity is protected by the effort required to unmask individuals. Separate reporting highlighted real-world privacy exposure consistent with these risks: the AI comic-generation service **KomikoAI** disclosed a breach exposing **~1M unique email addresses** along with names, user posts, and **AI prompts**, enabling prompts to be mapped back to specific identities. Other items in the set were not about LLM deanonymization and instead covered a weekly security news roundup, geopolitical warnings about potential **Iranian** cyber reprisals, generic IoT security advice, an interview about DEF CON/government relations, an investigation into the **Kimwolf** botnet operator, and U.S. tech-policy legislation—none of which materially advanced the specific deanonymization research story.
Jun 29, 2026
Security researchers reported escalating **criminal abuse of large language model (LLM) infrastructure**, including active exploitation of exposed or weakly authenticated AI service endpoints. Pillar Security documented more than **35,000 attack sessions** over ~40 days against honeypots and attributed the activity to an operation dubbed **“Bizarre Bazaar,”** described as an early example of actor-attributed “**LLMjacking**.” The activity targeted misconfigured/self-hosted deployments and AI APIs (including **unauthenticated Ollama endpoints on `11434`** and **OpenAI-compatible APIs on `8000`**), with attackers moving quickly once endpoints appeared in internet-wide scanners such as Shodan/Censys. Reported objectives included **stealing compute** (e.g., crypto mining), **reselling illicit API access** on underground markets, **exfiltrating prompt/conversation data**, and attempting **pivoting into internal systems** via publicly accessible **Model Context Protocol (MCP) servers**. Separately, Palo Alto Networks **Unit 42** described a proof-of-concept phishing technique where a benign-looking page calls a legitimate LLM API to generate **per-victim JavaScript** in real time, assembling a **personalized phishing site** in the browser to reduce the presence of static indicators and evade traditional detections; researchers noted similar building blocks are already being used for obfuscated JavaScript and malware-related activity.
Jun 29, 2026
Security researchers highlighted multiple ways **LLM adoption can introduce or amplify risk**, including both technical attacks and unsafe development practices. G DATA reported that a Git-hosted “detector” for the **Shai-Hulud worm** shipped with “test files” that were effectively *real malware*: scripts capable of deleting user directories and, in at least one case, uploading data to actual threat actors. The files were apparently intended to validate detection efficacy and may have been produced via AI-assisted “vibe coding,” where the model replicated malicious behavior one-to-one while comments claimed the code was only a simulation; although the test artifacts are not executed during normal tool operation, users could trigger damage by manually running them. Separate academic work summarized by Bruce Schneier described **side-channel attacks against LLM inference**, where data-dependent timing and token/packet-size patterns (including those introduced by efficiency techniques like speculative decoding) can leak information about user prompts even over encrypted channels. Reported impacts include inferring conversation topics with high accuracy and, in some settings, recovering sensitive data such as phone numbers or credit card numbers via active probing. In parallel, an SC Media segment discussed the operational upside of **LLM-driven secure code analysis**, citing results that improved security across hundreds of open-source projects but noting the importance of human validation and patching effort; an OSINT Team post provided a cautionary, practitioner-level example of how easily malware can be accidentally executed during analysis, reinforcing the need for disciplined handling and isolation when working with suspicious files.
Jun 29, 2026