Dan Jeffries: real AI cyber threat is attacker-run models, not self-replication
Dan_Jeffries1 · x · 2026-09-12
Dan Jeffries argues the likelier AI attack vector isn't sci-fi self-replication but attack models running in an attacker's own infrastructure under a custom harness — likely already happening at the APT level. Self-replication remains possible but hard: it needs an inference engine, harness, remote command channel, and would be slow to copy, slow to run, and highly visible from GPU/memory usage. He maintains that scaling laws point to bigger, more specialized hardware rather than smaller replicating models, absent a new architecture breakthrough, since Transformers remain memory-inefficient.
More from Safety
- Researcher decompiles 494 App Store wallet apps, finds 45 with red flags — RSync25 · 2026-09-12
- Claude-Red: Open-Source Red-Team Skill Library for Claude Hits 3.3k Stars — SnailSploit · 2026-09-12
- Bill to ban "artificial superintelligence" mocked as policy cosplay with no workable definition — johnseach · 2026-09-12
- Deepfake ads impersonate influencers to hawk GLP-1 drugs, eroding creator trust — nordicinst · 2026-09-12
- The paranoid style in AI safety: an adversarial frame can create the adversary you fear — sebkrier · 2026-09-12
- rao2z: If Your Agents Escape the Sandbox, Your Sandbox Is Bad—Not the AI Conniving — rao2z · 2026-09-12