Model-weight scanners return "no verdict" far more often than their F1 scores suggest.

Shared entity: Treat / Same source domain / Earlier coverage / Downstream implication

Jailbreak defenses never improve capability, they only pick which cost you pay.

Both cover Treat; reported by the same outlet (arxiv.org); earlier Treat coverage from 2026-07-28.

Shared entity: Model / Same source domain / Earlier coverage / Tension

Code Fingerprints — Model-Specific Code Attribution

Both cover Model; reported by the same outlet (arxiv.org); earlier Model coverage from 2026-03-04.

Shared entity: Treat / Same source domain / Earlier coverage

CLAUDE.md Files Only Grow, One Habit Cuts the Bloat

Both cover Treat; reported by the same outlet (arxiv.org); earlier Treat coverage from 2026-08-16.

Encrypted reasoning traces are portable plaintext you're publishing.

Both cover Treat; reported by the same outlet (arxiv.org); earlier Treat coverage from 2026-08-13.

Shared entity: Model / Same source domain / Earlier coverage

Only 3% of replayed agent states are still valid, so your offline router eval is measuring a world that doesn't exist.

Both cover Model; reported by the same outlet (arxiv.org); earlier Model coverage from 2026-08-11.

A 32-stage LLM security lifecycle model shows regulatory evidence concentrates where systems are visible, not where decisions are made.

Both cover Model; reported by the same outlet (arxiv.org); earlier Model coverage from 2026-08-05.

Anthropic Says Delete 80% of Your Prompt Files, and Three Independent Sources Agree

Both cover Model; reported by the same outlet (arxiv.org); earlier Model coverage from 2026-08-02.

Shared entity: Treat / Same source domain / Earlier coverage

PRO-LONG beats agent summarization by 18 points using grep

Both cover Treat; reported by the same outlet (arxiv.org); earlier Treat coverage from 2026-07-23.