AI Sec Watch

PraisonAI versions before 4.6.78 have a vulnerability where its defense against prompt injection (tricking an AI by hiding instructions in its input) is too weak. The defense only blocks attacks rated as CRITICAL threat level, but attackers can create single or double-vector attacks rated as HIGH threat level that slip through unblocked to reach the AI model.

This week's cybersecurity news includes a ransomware affiliate pleading guilty and agreeing to pay restitution, the discovery of QuimaRAT (a subscription-based remote access trojan, or malware that lets attackers control computers remotely) targeting multiple operating systems, Canada's intelligence agency disrupting ransomware operations, and a critical vulnerability in Writer AI that allowed attackers to bypass sandbox restrictions (security boundaries isolating code execution). Additional stories cover a trademark dispute between AI security companies, an exploit-selling startup exposed as a fraud operation, and a major insurance company data breach affecting 7 million people.

Security company CrowdStrike identified five new prompt injection techniques (attacks that trick AI systems into accepting harmful instructions hidden in normal input) that could threaten organizations using AI. These attacks include methods like hiding rules that activate later, bypassing safety features, delivering threats in stages, injecting fake control switches, and hiding malicious code in documents users upload to AI systems.

AI agents and other machine identities (automated accounts that aren't people) are creating a security gap because traditional identity security was designed around human behavior like hiring, role changes, and departures. Machine identities now outnumber human users by up to 50 to one in many organizations, but security teams often don't know what they are, who owns them, or what systems they can access, making them an easy entry point for attackers.

Researchers developed RFA-Tex, a method to create adversarial textures (specially designed patterns that trick AI systems) that can hide people from person detection AI at long distances. Previous adversarial textures only worked within 5 meters because fine details in the patterns broke down during long-range imaging, but RFA-Tex uses a new framework that preserves important details and reduces fragile structures, extending the attack range to 25-45 meters in real-world tests.

Researchers at Anthropic created a tool called the Jacobian lens to discover a hidden area in Claude (an LLM, or large language model) called J-space, which contains words related to responses the model is considering but may not ultimately generate. The article also covers various AI developments including OpenAI's new ChatGPT Work application and reports that OpenAI and Google sold AI models to blacklisted Chinese companies.

AI-powered surveillance systems combining facial recognition (technology that identifies people by analyzing their faces), real-time tracking, and mass databases are being deployed globally to monitor and immediately enforce rules, with China operating over 600 million AI cameras and the US Department of Homeland Security rapidly expanding its use for monitoring immigrants, protesters, and journalists. These systems create 'chilling effects' (where people self-censor and conform out of fear of being watched), raising concerns about discrimination, lack of transparency, and threats to democracy, as the automation of surveillance and enforcement removes human judgment and accountability.

Researchers have discovered 'HalluSquatting,' an attack that exploits AI hallucinations (when AI systems generate false information or invent things that don't exist) to deliver malware at scale. Attackers pre-register fake repository and package names that AI coding assistants commonly hallucinate, then plant malicious code in those fake repositories so that when users ask their AI tools to clone or install resources, the AI may pull down the attacker's code and execute it.

Deutsche Telekom, a major telecommunications company, is transforming itself into an AI-native organization by embedding AI (artificial intelligence) throughout its operations rather than simply adding it to existing processes. The company is using generative AI (AI that creates text, code, or other content) in customer service, network operations, and voice communications, with over 50,000 employees actively using ChatGPT Enterprise and seeing a 546% increase in AI tool usage since early 2026.