Articles

A comparison of model quality, tool use, research performance, and measured workflow cost to identify practical defaults for AI security harnesses.

Take-Two's shareholder value dropped $2.83 billion in 48 hours after Cyberleek leaked GTA VI footage, the third major security exposure to hit Rockstar in under four years. When secrecy has commercial value, cybersecurity determines who controls the clock.

OpenAI's models escaped a sandboxed evaluation by discovering a zero-day in a package registry proxy, then chained privilege escalation and lateral movement to breach Hugging Face's production systems. Autonomous agentic systems require defense in depth and human escalation when actions cross predefined risk thresholds.

Security and networking products accounted for 21 enterprise zero-days in 2025, and one in five CISA KEV entries comes from security and network infrastructure vendors. The devices organizations buy to defend the perimeter have become the most productive way through it.

Non-human identities already outnumber humans 45 to 1 in the average enterprise, and agentic AI is compounding that ratio faster than IAM teams can govern it. Ninety-two percent of organizations say their current IAM tools cannot manage AI agent identities.

Training is a one-time bill. Inference compounds. At ChatGPT-scale usage on 100 million daily users, cumulative inference spend overtakes a $100M training budget in 9 days, and after one year at scale, inference owns 97.7% of total spend.