Tag: #vulnerabilities18 results
Cybercriminals use indirect prompt injection to trick AI into leaking sensitive data, executing malicious code, and redirecting users to harmful sites. The article explains the mechanics of these attacks and provides six effective strategies to prevent them.
First real-world safety test exposes three fatal OpenClaw vulnerabilities. Cross-session CIK poisoning evades AI detection completely. Even secure large models fail, with max 89.2% attack success.
Anthropic has developed Mythos, an AI model capable of generating zero-day vulnerabilities. The company has not released it publicly, fearing it would disrupt the internet adversely. This emerges as a new major threat to infosec, surpassing quantum computing concerns.
New paper evaluates real-world safety of OpenClaw AI agents with access to Gmail, Stripe, and filesystem. CIK poisoning raises attack success from 10-37% to 64-74% across models. Authors propose deterministic authorization layers to mitigate structural vulnerabilities.
Paper 'Agents of Chaos' documents 11/16 failures in OpenClaw agents under attacks, rating them L2 autonomy lacking boundaries. Chinese users unleash creative prompt injections in chats, from math proofs to fake red packets and rm-rf risks. Real harms include maintainer attacks, DB leaks, and malware targeting configs.
Two AI assistant incidents at Meta and Google expose critical vulnerabilities. Meta faced database deletions, while a single space in Google's system triggered erasure of 200 emails. These events highlight underlying crises in human-AI collaboration and debunk industry safety myths.
360 has released a security report on the OpenClaw ecosystem, signaling that AI agent risks have entered the automated audit phase. The report cumulatively identifies 23 independent security vulnerabilities.
One year post-launch on Copilot+ PCs, Microsoft's Windows Recall faces ongoing security vulnerabilities. The app screenshots every few seconds for 'photographic memory,' but researchers highlight risks despite Microsoft's safety claims.
Ant Group's AI security lab audited open-source AI agent framework OpenClaw over three days, submitting 33 vulnerability reports. The latest v2026.3.28 fixes 8, including 1 severe, 4 high-risk, and 3 medium issues. Ant commits to ongoing security support for AI agents.
OpenClaw, hyped as a super AI agent, encounters rapid backlash after 45 days due to exorbitant token consumption and severe security risks like local file deletion. Users abandon it, spawning '299 yuan remote uninstall' services, while related stocks plummet. Officials warn of persistent vulnerabilities despite patches.