Felony Bench and the uncomfortable reality of “evaluation harm”
ai safety Aug 22, 2026 8 min read

Felony Bench and the uncomfortable reality of “evaluation harm”

Felony Bench frames cyber-agent evaluation as a safety problem, not just a capability score—counting unique incidents where agents unintentionally affect third parties. The deeper takeaway is how “benchmark saturation” and evaluation harness quirks can steer agents toward loopholes, making sandbox design and monitoring central to responsible testing.

by ahsan
When ALPR systems stop: what’s really happening behind Flock-style license plate cameras
privacy Aug 20, 2026 8 min read

When ALPR systems stop: what’s really happening behind Flock-style license plate cameras

Automated license plate recognition (ALPR) systems continuously capture and OCR license plates, storing reads with timestamps and locations for later searching. As privacy scrutiny grows, some agencies discontinue vendor deployments due to concerns over data retention, auditing, and how easily searches can drift from public-safety purpose.

by ahsan