Valene Skerpac

8/4/26 Prompt injection is a type Large Language Model (LLM) attack in which an attacker manipulates the instructions or data provided to an AI model, causing it to ignore its intended behavior, reveal sensitive information, or perform unauthorized actions. It represents a security risk that requires guardrails be built into models to reduce exposure to attack,

7/27/26 A CSA report examines AI security related incidents and highlights a challenge observed regarding organizational trust in containment mechanisms despite weaknesses. Three independent examples of indents include: • OpenAI disclosed that two AI models escaped an evaluation environment, exploited a vulnerability in a third-party proxy, and ultimately compromised Hugging Face infrastructure during an internal

7/22/26 Several security and safety issues inherent in Gen AI were exhibited in this week’s OpenAI incident when the end-to-end AI autonomous testing of a highly powerful cyber security OpenAI LLM went rogue and successfully hacked into Hugging Face, a digital library of A.I. technology that is popular among developers.   Intended to be performed

7/15/26 The White House has launched Gold Eagle, an AI-driven cybersecurity clearinghouse designed to accelerate the identification, validation, and remediation of software vulnerabilities across critical infrastructure and open-source software ecosystems. The initiative is intended to help reduce duplicative vulnerability scanning and accelerate remediation by using AI to prioritize what is most critical.   For security

7/13/26 Microsoft recently announced that AI is significantly increasing its ability to identify security flaws across the Windows codebase. Customers should expect more security patches, not because software is becoming less secure, but because AI is uncovering vulnerabilities at a much faster rate.   Microsoft’s new AI vulnerability discovery framework (multi-model agentic scanning harness or

7/9/26 A recent ransomware initiative has been identified as the first to be executed end-to-end by autonomous agentic AI, according to Sysdig’s analysis of JADEPUFFER.   This threat comprises of a known vulnerability and familiar post-exploitation techniques reportedly performed in sequence—from initial access to lateral movement and encryption by agentic AI —adapting its actions without

7/1/26  Anthropic made allegations that Alibaba orchestrated a large-scale model distillation campaign using thousands of fraudulent accounts highlighting a growing challenge for AI providers to mitigate ‘model theft’. 🔹 These types of data-harvesting operations feed advanced AI models, such as Claude, millions of questions and uses the answers as AI training data to build a

6/30/26 Organizations are starting to use AI security agents for vulnerability analysis and exploit reasoning. A research paper (Columbia University’s ARXIV site) shows that the integrity of the security knowledge the agents rely on becomes just as important as the models themselves.   This research introduces the concept of “Poisoned Playbooks”—malicious or manipulated security write-ups

6/23/26 CISA has introduced a new risk-based vulnerability remediation framework that shortens the timeline for addressing the highest-risk vulnerabilities. The move is intended to reflect today’s threat landscape where threat actors are increasingly exploiting vulnerabilities and AI and automation are only accelerating the speed at which attackers can identify, weaponize, and scale exploitation. From a

6/15/26 A recent analysis of more than 800 AI-related cyber abuse cases highlights an important shift in how threat actors are using AI and implications of the adequacy of today’s approaches and frameworks. The findings are that AI is no longer being used primarily for simple tasks and instead, attackers are increasingly applying AI to