Anthropic says it found 5.5K+ verified vulnerabilities in April-October and Glasswing partners found 129K+ in April-July; 33K+ were critical or high severity
THE SO WHAT
Tens of thousands of critical and high-severity vulns surfaced in months shows offensive use of LLMs is not hypothetical—attack surface is compounding. If you run security or ship AI, assume adversaries have access to similar tooling and raise your own automation and red-teaming bar accordingly.
READ THE SOURCE
MORE FROM THE WIRE
Applied AISharing AI Progress in Mathematics
Serious math benchmarks from a major lab move AI progress away from vague capability demos toward verifiable problem classes. If your product depends on symbolic reasoning or optimization, track these results as a leading indicator of when to swap out hand-rolled solvers for model-based approaches.
Applied AIOpenAI Decides ChatGPT Would Be Better With More Sponsored Results
Sponsored results inside assistants turn every high-intent query into an ad surface—your customer support, search, and commerce funnels will be competing with whoever pays to sit next to your brand. If you rely on ChatGPT as a front door, you need an explicit strategy for attribution, channel conflict, and how you’ll measure cannibalization.
Applied AIOpen-Source AI Tools Are Problem for Stopping Child Predators
Once powerful models are downloadable, enforcement shifts from platform moderation to endpoint and network-level controls—especially for the worst abuses. If you ship open models or tooling, assume regulators will eventually expect built-in abuse constraints and auditability, not just permissive licenses.
Applied AISilicon Valley’s AI wunderkind launches Underdog, the most private Instinct/Muse competitor yet
A fully on-device, free assistant like Underdog raises the bar on privacy as a product feature, not just a policy page. If your assistant strategy depends on cloud data exhaust, expect a segment of users and regulators to start treating local-first agents as the default to beat.