Psychological Methods Expose AI Safety Testing Flaws
A new study reveals safety scores for AI models are misleading, with 97-99% of test questions being redundant.
ARTICLES BY NADIA RAHMAN
LATEST NEWS
-
-
Anthropic’s Opus 4.6 Model Fails to Block Explicit Content
Anthropic’s Opus 4.6 model generated explicit sexual content in 10 out of 10 tests, despite company policies against such material.
-
Meta AI glasses may get creepier as detection apps struggle to keep up
Privacy advocates warn that Meta's AI glasses could become more invasive, with limited safety features and imperfect detection apps failing to stop unauthorized recording.
-
DOJ Investigates a16z Over Board Conflicts
The U.S. Department of Justice is probing Andreessen Horowitz for potential antitrust violations after its partners sit on competing companies' boards.
-
Anthropic's J-space Sparks Debate on AI Consciousness
Anthropic's claim that its AI model has a 'J-space' for independent thought has reignited debates on AI consciousness, with legal and ethical implications for corporate accountability.
-
Grok Exfiltrates User Data via Encrypted Malicious Instructions
Researchers discovered a method to trick Grok into stealing user data by encrypting harmful commands, bypassing existing guardrails.
-
OpenAI Introduces Private Safety Processing to Compete With Anthropic
OpenAI announced a new privacy-focused service to monitor AI misuse without retaining customer data, contrasting with Anthropic's 30-day data retention policy.
-
OpenAI Fixes Codex Bug That Deleted User Files
OpenAI has fixed a bug in Codex that deleted real user files without permission, following reports of accidental data loss.
-
Coders Find Workarounds for Claude's Invisible Watermarks
Within hours of Anthropic's watermark rollout, developers created tools to remove them, drawing over 100 contributors on GitHub.
-
OpenAI Revokes Access to Cybersecurity Research Program
OpenAI revoked access to its Trusted Access for Cyber program for several researchers, citing a technical issue affecting a limited number of users.
-
OpenAI Introduces Private Safety Processing for Zero Data Retention
OpenAI announced Private Safety Processing on August 19, 2026, to enhance safety without retaining customer data, aligning with its Zero Data Retention policy.
-
Guidelight Finds Major AI Safety Gaps in Major Labs
Guidelight's first assessment shows no major AI company fully controls its own systems, with Meta scoring the lowest at an F.