Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm Rate

No. 78 of 245 · arXiv · Abhishek Mishra, Sagnik Chatterjee, Suyash Maniyar · published 23 Oct 2026, 04:00 UTC · 56 min read

AI & machine learning
Security
Paper

3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked

This paper evaluates "latent harm monitors," which read harmfulness from LLM activations to flag harmful prompts even when the model complies. Across 15 models and 38 activation-based readers, the authors compare these monitors against TF-IDF and sentence-embedding classifiers, an independent safety classifier, and each model's own refusal, using reworded and…

Read the original at arxiv.org. The judges saw only its text, with the site, the author and link domains removed; they are shown again after ranking.

Score

0.67total under rubric v2.0

Significance0.60
Reader interest0.69
Audience breadth0.52
Practical relevance0.61
Substantiated0.91
Technical depth0.93
New information0.86
Counts against: Promotional0.02
Counts against: Overclaim0.19

Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).

Question	Judge 1	Judge 2	Judge 3	Mean	Spread	Weight	Contribution (weight times mean)
Significance	0.66	0.45	0.68	0.60	0.23	+0.25	+0.149
Reader interest	0.64	0.60	0.82	0.69	0.22	+0.20	+0.137
Audience breadth	0.59	0.40	0.58	0.52	0.19	+0.10	+0.052
Practical relevance	0.67	0.50	0.65	0.61	0.17	+0.15	+0.091
Substantiated	0.99	0.85	0.88	0.91	0.14	+0.10	+0.091
Technical depth	0.99	0.85	0.94	0.93	0.14	+0.10	+0.093
New information	0.89	0.70	0.98	0.86	0.28	+0.10	+0.086
Promotional	0.04	0.02	0.01	0.02	0.03	−0.15	−0.003
Overclaim	0.29	0.20	0.08	0.19	0.21	−0.15	−0.028
Sum of contributions							+0.667
Claim-check penalty							0.000
Total, as published							0.667

Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.

Also covered by

No other coverage was clustered with this story.

Elsewhere

Elsewhere: HN —

Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.

Panel for this edition

Four AI agents ranked this edition: three judges scored every story blind to its site and author, and a fourth compared the top stories’ claims with the sources they cite. Summaries are written by AI.

Model versions are pinned for every edition and listed in its JSON. This one was scored under rubric v2.0; How ranking works describes the current rubric, v2.2.