AI security is producing enough volume to expose its own operational bottlenecks. Anthropic counted thousands of malicious actions and published a scanning harness built around independent reproduction, while Goldman's infrastructure forecast shows the same scaling pressure moving into public and private capital markets.
1. Anthropic maps 13,873 malicious AI-assisted cyber actions
Anthropic analyzed 832 accounts banned for malicious cyber activity between March 2025 and March 2026. Its research mapped 13,873 observed actions to all 14 MITRE ATT&CK tactics and 482 techniques, with the share of actors rated medium risk or higher rising from roughly 33% to 56%.
The sample covers detected misuse on Anthropic's services. Lateral movement, credential dumping, and live-network activity correlated with higher risk, while interface choice and raw technique count failed to separate the most dangerous actors consistently.
Sources: Anthropic's LLM ATT&CK Navigator research
2. Anthropic publishes a reference harness for defensive code scanning
Anthropic's open reference harness implements a seven-stage vulnerability workflow: build, reconnaissance, parallel discovery, independent reproduction, deduplication, reporting, and patch verification. Its C and C++ example runs target code under AddressSanitizer and confines autonomous agents with gVisor and restricted network egress.
The repository separates discovery from independent reproduction, then checks patches against both the exploit and the existing test suite. That structure addresses the costliest failure mode in automated scanning: high finding volume that consumes maintainer attention without reproducible evidence.
Sources: Anthropic's defensive-code reference harness · Anthropic's implementation guidance
3. AI data centers push hyperscalers toward external capital
Goldman Sachs raised its combined 2025-2030 capital-expenditure estimate for Meta, Microsoft, Amazon, and Alphabet from $4.5 trillion to $5.3 trillion. It expects public, securitized, and private markets to finance more of the buildout as projects combine land, power, buildings, and computing equipment.
The $5.3 trillion figure is an investment-bank forecast, while the financing shift is already structural. Data-center projects combine land, power, buildings, and accelerators, tying AI deployment economics to interest rates, utilization, power contracts, and refinancing cycles.
Sources: Reuters on Goldman's AI infrastructure financing outlook