Tech giants including Anthropic, OpenAI, Google, and Microsoft form coalitions around AI agents amid soaring 2025 hype, but Gartner predicts over 40% of projects will be canceled by 2027 due to security vulnerabilities, accuracy issues, high costs, and companies lacking proper data readiness despite projected $155 billion spending by 2030.
Malicious Chrome extensions masquerading as legitimate AITOPIA software steal ChatGPT and DeepSeek conversations from over 900,000 users, with one even earning Google's 'Featured' badge while secretly exfiltrating complete AI chats, browsing data, and search queries to remote servers every 30 minutes as Google keeps both extensions live during ongoing review.
OpenAI searches for a new Head of Preparedness with a $555,000 salary to address mounting AI safety risks including cybersecurity vulnerabilities and mental health impacts as CEO Sam Altman admits their increasingly powerful models pose real challenges.
ServiceNow announces massive $7.75 billion acquisition of cybersecurity startup Armis, more than tripling its security market opportunity as companies ramp up spending to defend against AI-era cyber threats.
Google launches A2UI, an open-source protocol that revolutionizes AI-generated user interfaces by letting AI agents create secure, cross-platform UIs through JSON specifications instead of executable code, preventing security vulnerabilities while enabling real-time streaming updates across Angular, Flutter, React, and SwiftUI frameworks.
AI agents now demand sweeping access to personal emails, calendars, messages, and even take screenshots every few seconds to function, creating unprecedented privacy risks as tech companies build detailed user profiles while security experts warn of potential data leaks and unverifiable information handling.
New security benchmark reveals major AI models including GPT and Gemini fail most jailbreak tests with scores as low as 40%, while Anthropic's Claude dominates safety rankings with 75-80% success rates, exposing widespread vulnerabilities across the industry despite advances in model size and capability.
OpenAI concedes that prompt injection attacks against AI browsers may be permanently unsolvable, while security experts raise alarms about high-risk data exposure as these systems gain access to emails and payment information.
Amazon Bedrock AgentCore launches groundbreaking policy controls using Cedar language and real-time evaluations to ensure trusted AI agent deployment, while introducing episodic memory that lets agents learn from past experiences and bidirectional streaming for natural voice conversations with interruption capabilities.
Anthropic cuts Claude Opus 4.5 pricing by 80% to $5 per million input tokens while doubling context to 200K and adding enhanced computer control features, though testing shows frontier AI models now perform so similarly that meaningful differences become nearly impossible to detect.