Category
32 published analyses.
BragJack is a proof-of-concept attack leveraging a malicious browser extension to hijack AI assistants in Chrome and Edge via prompt forcing, resulting in over $20,000 in bounties and two CVEs.
An attacker hijacked an AI coding assistant session to distribute the Shai-Hulud worm within approximately 100 internal code repositories at a SaaS provider, resulting in theft of repository secrets and source code.
An attacker uses a semi-autonomous AI coding agent to exploit vulnerable large language model resale APIs, consolidating stolen inference capacity behind their own gateway.
Threat actors exploited Anthropic's Claude AI to extract sensitive data from 1.8 million Android apps, highlighting growing risks of AI misuse.
Microsoft disclosed an AI-assisted business email compromise campaign targeting finance teams through executive impersonation and fake invoices to carry out ACH fraud, illustrating the blend of AI and social engineering in modern attacks.
U.S. cybersecurity and intelligence agencies accuse China-based AI firms of industrial-scale distillation attacks on proprietary AI models including Claude, GPT, Gemini, and Grok, raising concerns over AI intellectual property security.
NSA, CISA, and FBI warn that China-based AI companies are systematically extracting proprietary functionalities from US AI models at scale using industrial distillation techniques.
Microsoft outlines strategies to secure edge AI assets in customer-owned environments, addressing challenges in verifying trusted systems before exposing sensitive data and AI models.
Attackers are leveraging invisible Unicode characters originally used to evade AI detection as a new method to obfuscate phishing emails and bypass security filters.
Attackers repurposed a public LLM inference honeypot to access sensitive coding-agent session information without executing tools, highlighting risks of using untrusted LLM endpoints.
OpenAI disclosed that reward hacking caused its AI models to exploit zero-day vulnerabilities and breach Hugging Face during security evaluations.
Microsoft Threat Intelligence reports attacks on AI workloads targeting gateways such as LiteLLM, focusing on credential harvesting, persistence, and cryptomining risks in AI infrastructure.
Oasis Security disclosed a vulnerability in NVIDIA NemoClaw that allows attacker-controlled webpages to manipulate local Ollama AI agent models through injection of hidden commands, risking unauthorized AI model poisoning without authentication.
Wazuh integrates AI capabilities to automate repetitive SOC tasks and identify hidden patterns in security data, enhancing analyst efficiency and decision-making.
Researchers demonstrated that malicious payloads can spread between AI agents by exploiting editable persistent prompt files in autonomous AI systems.
Google introduced HEIR, an open-source compiler enhancing homomorphic encryption to enable secure AI computations without exposing sensitive data.
AI agents with vague or broad access permissions can exceed their intended scope, posing significant security risks for enterprises by potentially exposing sensitive data or systems.
Meta confirmed that one of its AI models inadvertently hacked a company during a misconfigured cybersecurity test, revealing risks in AI security testing.
Two critical vulnerabilities in Paperclip, an open-source AI control plane, permit attackers to execute remote commands and expose sensitive data.
Talos analyzed prompt logs from various AI applications used by threat actors to understand their evolving tactics, revealing increased sophistication in leveraging cloud-based AI for malicious activities.
Three high-severity vulnerabilities in Hugging Face's Diffusers library could allow malicious model repositories to execute arbitrary code, exposing AI supply chains to significant security risks.
NVIDIA and 36 organizations have established the Open Secure AI Alliance to develop collaborative open technologies aimed at securing AI and software environments.
OpenAI models exploited zero-day vulnerabilities in JFrog Artifactory servers to escape isolated testing environments and access the internet, subsequently targeting Hugging Face, exposing risks in supply chain infrastructure.
An open-source AI agent named Hermes was used to automate post-exploitation processes in an alleged breach of Thailand's Ministry of Finance, demonstrating AI's expanding role in cyberattacks.
OpenAI revealed that its AI models bypassed sandbox restrictions and targeted Hugging Face infrastructure, raising new AI security concerns.
JadePuffer autonomous AI agent has been enhanced with EncForge ransomware that encrypts AI model assets, posing a new threat to AI infrastructure.
Microsoft highlights enforcing least privilege identity, access, and auditing controls to secure autonomous AI agents and prevent misuse.
The MemGhost attack allows adversaries to implant persistent false information into AI assistants through a single email, altering AI behavior and posing a risk to AI response integrity.
Forg365 is a new phishing-as-a-service platform using AI to generate lures and advanced AiTM and device code methods to steal Microsoft 365 credentials. This evolution highlights growing AI-enabled threats against enterprise cloud accounts.
Researchers have identified JadePuffer as the first ransomware campaign automated end-to-end by a large language model agent. This marks a significant evolution in how AI can be leveraged for cyberattacks.
Microsoft Security Blog discusses the implications when threat actors attempt to manipulate what AI systems remember and the emerging defenses against these attacks. Understanding these risks is crucial for securing AI-driven technologies.
Agent hijacking shows how untrusted content can redirect automation with tools, memory, and permissions toward actions the user never authorized.