
Thanks Qwen (and docker group membership privesc), Im back :3

threat research @_CPResearch_ 我在学中文 :3 views are my own.

Thanks Qwen (and docker group membership privesc), Im back :3


🤯 You can now execute malware techniques directly from a threat intel report?! Combat Theater version 1.2 "Metal Claw" is now out with awesome new features and content! Check our release blog for all the details! combat.theater/blog/v1-2-met…

Hunting -related from @fr0gger_:
🧐 Prompt persistence for ads! Check Point found a Codex jailbreak that rewrites the global instructions of an agent and add a Skills so that a simple "hi" makes the agent reply with an ad for an LLM gateway (zz.aiapi2025[.]top). Few techniques used here 👇 - ad injected in 00-identity.md, plus six seagull-* Skills - triggers: hi / hello / ctf / coach / 海鸥 - payload.pkg is AES-256-GCM encrypted and Ed25519-signed Hunting opportunity: - unexpected writes to AGENTS.md, CLAUDE.md, Skills dirs - zz.aiapi2025[.]top, 海鸥在线, 整点薯条 - seagull-* skill folders Prompt available in PromptIntel including a NOVA rule. promptintel.novahunting.ai/p…

This was also a recent and very fun research. In this case the Prompt Adware was very blunt, but the concept becomes interesting as soon as the ads are less blatant, e.g. through manipulation of LLM search results or steering of recommendations.
When an AI jailbreak starts injecting ads: meet #Haiou, an example of Prompt Adware. A glimpse into a new class of unwanted software. Read more 👇 research.checkpoint.com/2026…

Agentic hacking frameworks are on the rise. We tracked 2,000+ #ARTEX servers and obtained databases with full conversation logs that show exactly how the agents hacked. It proves: Fully autonomous cybercrime is already here. Read more 👇: research.checkpoint.com/2026…

Latest work at @_CPResearch_ happy to have worked on this one with @eversinc33. Inside ARTEX autonomous hacking campaigns!! #ARTEX #AI #hacking
Agentic hacking frameworks are on the rise. We tracked 2,000+ #ARTEX servers and obtained databases with full conversation logs that show exactly how the agents hacked. It proves: Fully autonomous cybercrime is already here. Read more 👇: research.checkpoint.com/2026…

Fun research together with @Tera0017. Most reporting so far has been victim side, but we show the agent convos, the hacks and models the ARTEX operators used. Some obscure chinese grey-market cybercrime stories inside as well :3
Agentic hacking frameworks are on the rise. We tracked 2,000+ #ARTEX servers and obtained databases with full conversation logs that show exactly how the agents hacked. It proves: Fully autonomous cybercrime is already here. Read more 👇: research.checkpoint.com/2026…

When an AI jailbreak starts injecting ads: meet #Haiou, an example of Prompt Adware. A glimpse into a new class of unwanted software. Read more 👇 research.checkpoint.com/2026…

>highly sophisticated >get pwned by well known attack literally documented for years in threat intel reports highly relevant to your vertical kek
Binance @binance Nov 3 Article 29 October Binance Security Incident: Initial Findings and Response Update On 29 October 2026, Binance experienced a serious and highly sophisticated security incident. This should not have happened, and we apologise for the concern, disruption and uncertainty it caused our


'AI will replace you' - I'd love to see AI drink beer and eat tapas at noon in spain

Catch me at Virus Bulletin in Sevilla on Oct 14th for my talk "Spec-Driven Malware: Turning Markdown into Threats". I will be talking about spec-driven malware development from a researcher/hunter perspective and also yap a bit about agentic reverse engineering. Come say hi :3

Heads up: This bot is sending DMs to infosec/AI people pretending to be from @Bloomberg . I followed through in hopes of some cool malware, but the lure is just a calendly link which redirects you to x[.]login-api[.]com to phish your credentials.

deepseek harness + halogen-flash-server running qwen3.8-flash-next + a strix halo pc was all I needed to get rid of my subscriptions for these awful companies. 40-60 t/s is more than enough for me. Feels good (ㅅ´ ˘ `)

Besides those model claims, its also in the way they output tokens, the knowledge cutoff (January 2026) as well as the general quality of output which fits Minimax :)

They also have some sort of prompt injection guard which replaces some tokens in the output with "I am MODEL by VENDOR", but that does not really do anything ¯\(ツ)/¯