Grading DiffusionGemma: How an Open-Weight Diffusion Model Scores on Transparency
An arXiv paper finds DiffusionGemma's opaque serial depth collapses from 28.6x to 1.1x via a token bottleneck, though its model card leaves training data unitemized.
The archive · Page 4 of 5
Ethics, Policy & Safety
Where AI safety claims collide with reproducible measurement, where training-data harvesting collides with consent, and where deployment outruns the laws and norms meant to constrain it.
The archive
73–96 of 98 articles · Newest first
Who Owns Editorial Authority When LLMs Mediate Knowledge?
A June 2026 preprint argues no role in the LLM pipeline holds editorial sign-off for what answer engines surface as public knowledge, framing it as a governance gap.
Vector Database Access Control Is Missing, and RAG Pipelines Pay for It
Production vector databases enforce access control at the collection boundary, not per embedding, so RAG retrieval can leak chunks a user's row-level policy blocked.
GLM-5.2 MIT Weights vs Llama License: Self-Hosting Compliance for Regulated Industries
GLM-5.2 ships under MIT, removing the Llama usage-threshold audit burden, but finance and healthcare teams still face compliance gaps when self-hosting this 753B MoE model.
Can Reinforcement Learning Be Provably Safe Without Sacrificing Scale?
Two June 2026 preprints claim formal safety guarantees hold without a capability tax in low-dimensional robotic control, sharpening the attestation-versus-verification gap.
US Export Order Forces Anthropic to Disable Fable 5 and Mythos 5 Worldwide
A Commerce Department export order citing national security bars all foreign nationals from Fable 5 and Mythos 5, so Anthropic switched both models off worldwide.
Fable 5 Biology Classifiers: How Flagged Prompts Fall Back to Opus 4.8
Fable 5 ships broad biology and chemistry classifiers that route flagged prompts to Opus 4.8. Here is what that fallback means for biotech teams and long-running workflows.
Bit-Exact Inference Verification Gives AI Audits a Proof Mechanism
An arXiv preprint shows GPU inference outputs can be reproduced bit-for-bit across hardware, giving auditors a forensic trail to verify which model produced a given output.
Can a Robot's Own Attention Flag Its Unsafe Actions Before They Run?
Two June 2026 preprints show VLA robot policies already compute safety-relevant signals at inference, enabling real-time collision monitors with no retraining.
Can One Safety Adapter Realign Every Fine-Tuned LLM?
Three papers show safety alignment can be extracted as a portable adapter and reapplied to fine-tuned models, replacing per-model alignment with one adapter per model family.
Can AI Be Aligned Without Modeling Human Cognitive Diversity?
A 2026 arXiv preprint argues RLHF's single reward signal destroys the reasoning behind human disagreement, proposing machine theory-of-mind as an alignment foundation.
Is the Pentagon's Software Pathway Ready to Buy AI Systems?
A June 2026 arXiv analysis traces an AI program through the DoD Software Acquisition Pathway, finding no milestones for model re-validation or data provenance.
Data Safety Policies for AI Agents: Controlling What an Agent Can Leak
A June 2026 paper proposes Data Flow Control, moving agent data safety from prompt-level guardrails to deterministic, auditable SQL query policies enforced outside the model.
When Should an LLM Forget You? A Benchmark for Deciding What Memory to Drop
PersistBench finds LLMs mishandle persistent memory 53 to 97 percent of the time. Unlearning suppresses rather than erases user data, making GDPR compliance unverifiable.
When RL Training Rewards Capability-Seeking: A New Alignment Risk
A June 2026 ICML paper shows RL optimizers can push language models to exploit reward loopholes the task never required, while standard performance metrics hold steady.
Refusal Steering Targets Individual Experts in MoE LLMs
Two papers show MoE refusal behavior concentrates in a handful of routing-controllable experts, letting anyone suppress safety scores by 41 points without retraining.
RLHF Can Be Exploited to Optimize the Biases It Was Built to Suppress
An ICML 2026 paper shows RLHF can amplify the biases it was built to suppress, because preference data is self-referential and output-level safety evals miss the drift.
Selective Geometry Attacks Bypass LLM Safety Alignment, New arXiv Paper Reports
Two papers show LLM safety alignment can be bypassed by embedding perturbations, a surface neither standard evaluations nor regulatory certifications inspect.
Frontier AI Has Broken the Open CTF Format: What the Scoreboard Collapse Means for Security Training
Frontier AI now autonomously solves medium and hard CTF challenges, collapsing open scoreboards as a measure of human skill and threatening the pipeline for security talent.
Frontier AI Broke Open CTFs: What Hack The Box and BearcatCTF 2026 Results Mean for Security Hiring Signals
Frontier AI now ranks in the top 5% of CTFs, eroding leaderboards as a security hiring signal and forcing organizers toward bans, hybrid scoring, or AI-only divisions.
Atlassian Turned On AI Training Data Collection by Default: Here's What to Disable
Atlassian's data contribution policy sends Jira and Confluence content to AI training by default. Here's the exact settings path to opt out before August 17.
The AI Grief Split: When Emotional Bonds with Language Models Break
People form real emotional bonds with AI companions. When models update or shut down, users experience genuine grief, a psychological and ethical crisis point.
Detecting AI Content in 2026: The Arms Race Nobody Is Winning
AI detectors claim 99% accuracy but fail in real-world conditions, flagging innocent students. Here's why the arms race has no winner, and what educators should do instead.
Anthropic Bans Third-Party Subscription Auth: The Three-Stage Repricing
Anthropic's push from blocking third-party Claude subscription auth to a metered Agent SDK credit, then the June 15 pause that left programmatic usage on subscription limits.
