Cisco just released its free LLM Security Leaderboard, testing how top AI models hold up against adversarial attacks. Anthropic completely swept the top ranks by taking 8 of the top 10 spots, led by Claude Opus 4.5 at No. 1, while OpenAI’s GPT-5.2 finished 7th. Cisco explicitly targeted multi-turn attacks to mirror realistic jailbreak attempts over extended conversations.
Anthropic just analyzed millions of interactions across Claude Code and its API to see how agent autonomy works in practice. The longest Claude Code sessions nearly doubled to over 45 minutes of autonomous work, while experienced users auto-approve actions over 40% of the time. Why it matters: Models can handle far more autonomy than we currently give them, and experienced humans are shifting from micromanaging to supervising from afar.
Anthropic just released Claude Opus 4.5, calling it the world's best model for coding, agents, and computer use. It even outperformed human candidates on the company's internal software engineering exam. With API pricing lowered to $5/$25 per million tokens and a new effort parameter to cut token usage, frontier intelligence is getting both smarter and cheaper.
Anthropic just rolled out new defenses against prompt injection attacks, helping Claude Opus 4.5 resist malicious instructions hidden across the web. The upgrades combine reinforcement learning with improved classifiers to drive attack success rates down to 1%, allowing the Claude for Chrome extension to expand into beta for Max plan subscribers. Why it matters: as agents take over real-world browsing and form filling, securing untrusted web content is critical—though Anthropic stresses a 1% risk is far from a solved problem.
