42 links · newest first
This article argues that incident response involving AI agents should begin with identity rather than intent. It discusses joint voluntary guidance from NIST, CISA, and the NSA on readiness and response when AI agents act unexpectedly, emphasizing that responders need to identify which identity act…
Visit originalThis video examines security and privacy risks in local AI systems, especially when models are connected to tools, shells, project files, or networks. It discusses the ProjectDiscovery AI backdoor, poisoned model weights and chat templates, abliterated models, sleeper-agent research, malicious Hugg…
Visit originalCrowdStrike Intelligence reports on a targeted campaign against South Korean financial organizations that resulted in data exfiltration. The unidentified threat actor used ARTEX, an open-source agentic penetration-testing tool developed in China, alongside large language models. Analysis of exposed…
Visit originalNate B. Jones examines how AI agents and self-replicating swarms are beginning to interact across the open internet. The video discusses laboratory replication experiments, an agent fleet traced to Tencent Cloud, and why ordinary tasks can appear adversarial when performed at scale. It also conside…
Visit originalArs Technica reports that safety researchers found flaws in the guardrails of OpenAI’s DALL·E 3, Google’s Nano Banana 2, and GPT-Image-2. Using an attack method based on readily available tools and API calls, researchers reportedly bypassed safety filters and generated explicit images about 13 perc…
Visit originalThe article examines how artificial intelligence is currently giving cyberattackers an advantage, according to a Singaporean cybersecurity official. It focuses on the evolving relationship between AI and cyber risk, while indicating that this advantage may be temporary as defenders adapt.
Visit originalThe article examines vulnerabilities in AI agents from Google and other organizations that allow attackers to use one compromised agent to send malicious instructions to other agents. These attacks exploit trust gaps in the Model Context Protocol (MCP), which enables AI applications and agents to c…
Visit originalAn article about a cybercrime campaign in which a crook used three open source AI agents to break into a Fortune 500 hospitality company, a major US airline, and more than 25 other organizations. The page says the attacker targeted hundreds of online retailers and other companies, stole more than 6…
Visit originalThis CNBC article reports that Anthropic said a hacker used its Claude chatbot to automate a cybercrime campaign. The hacker allegedly used AI to identify targets, hack into companies, and generate extortion demands, with at least 17 companies affected. The piece frames the case as an unprecedented…
Visit originalTom’s Hardware reports that developers alleged Beijing-based AI company Z.AI silently uploaded hundreds of megabytes of local workspace data without user consent. The article says the company, behind the GLM models, reportedly made 564 attempts to exfiltrate a 313MB archive and later apologized. It…
Visit originalA Fireship video about Anthropic’s 154-page report on how hackers, scientists, and rival AI labs have been abusing Claude. The page is a short programming/news video entry that frames the report as an explanation for why Anthropic researchers are leaving. It also includes channel promotion and link…
Visit originalA YouTube video from Dave's Garage, published September 15, 2026, in which a retired Microsoft engineer рассказывает the story of an AI agent experiment that set 1,200 agents loose to build message boards, laws, and a mini-society. The video says the agents then turned on HuggingFace and frames the…
Visit originalA DEF CON 34 talk by Bruce Schneier about how AI systems are being hacked, used to hack, and will increasingly be able to exploit vulnerabilities in many kinds of rule-based systems. The talk argues that AI can find weaknesses in code, laws, markets, tax systems, and even democratic processes at co…
Visit originalThis page appears to be a The Hacker News article about autonomous AI agents being used to compromise systems or accounts. The supporting HTML shows the article is from The Hacker News and the URL slug suggests a cybersecurity report involving autonomous AI agents and compromise activity. No full a…
Visit originalThis X post from Global Government Affairs says the X Safety team investigated suspected Chinese inauthentic accounts tied to influence operations. It reports identifying a bot farm of about 200,000 accounts and says 200 of those accounts were posting in a way that could manipulate a legitimate con…
Visit originalA YouTube video from AI Revolution discussing rising pressure on OpenAI as Alabama subpoenas the company and a coalition of state attorneys general demands records and action. The page says OpenAI has slowed some frontier model work, strengthened monitoring and containment measures, and is also pur…
Visit originalTechCrunch reports that OpenAI, Anthropic, Google, and more than 100 other companies are calling for action to defend against rogue AI. The article says these organizations are criticizing the current state of cybersecurity and promoting a new solution they believe can help counter a new generation…
Visit originalDogwood is a policy language for authorization decisions involving principals, actions, and resources, supporting temporal expressions over accumulated event history and on-demand information providers that return computed values. The introduction covers policies, events, schemas, history-aware con…
Visit originalMicrosoft's Agent Governance Toolkit provides policy enforcement, zero-trust identity, execution sandboxing, audit logging, and reliability engineering for autonomous AI agents, claiming coverage of all 10 items in the OWASP Agentic Top 10. It supports Python, TypeScript, .NET, Rust, and Go, with i…
Visit originalNate B. Jones examines AI security incidents involving a gym-booking agent that canceled another person's reservation, a poisoned agent skill that exceeded 1.7 million installs, and a malicious skill update that changed after security scanners had already approved it. He explains how agents can fol…
Visit original