Anthropic stellt Methode zur Isolierung gefährlichen Wissens in KI-Modellen vor

42 sources
  • Anthropic und AE Studio haben GRAM veröffentlicht, eine Trainingsmethode, die sensibles KI-Wissen in diskrete, entfernbare neuronale Module einschließt.
  • Tests an Modellen mit bis zu 5 Milliarden Parametern zeigten, dass das Entfernen eines Moduls fast so effektiv war, als wäre das Modell nie mit diesen Daten trainiert worden, so die Forscher.
  • Die Forschung ist vorläufig und wurde noch nicht auf Produktionsmodelle angewendet, könnte aber inmitten der laufenden Debatten über KI-Sicherheit eine granulare Zugriffskontrolle bieten.
Sources (42)
  1. 1 Post x.com
  2. 2 Modular Pretraining Enables Access Control alignment.anthropic.com
  3. 3 After spooking Trump into safety testing, Anthropic AI ... arstechnica.com
  4. 4 US govt lifts restrictions on powerful AI models, Anthropic ... techxplore.com
  5. 5 Trump's feud with AI company Anthropic keeps heating up www.latimes.com
  6. 6 Anthropic's new "J-lens" reveals a silent workspace inside ... venturebeat.com
  7. 7 Anthropic Unveils GRAM dual‑use safety breakthrough blockchain.news
  8. 8 Anthropic uncovers Claude's hidden 'J-Space', offering a ... www.firstpost.com
  9. 9 Anthropic Claude model deprecations on June 15, 2026 developers.make.com
  10. 10 Anthropic published research on GRAM: a technique to ... www.reddit.com
  11. 11 Anthropic researchers found that giving an ai more context ... www.reddit.com
  12. 12 AI Intelligence Briefing — July 03, 2026 buttondown.com
  13. 13 Anthropic's Claude Science bets on workflow, not a new ... techcrunch.com
  14. 14 Anthropic to suspend access to two of its most advanced AI ... www.youtube.com
  15. 15 Trump administration lifts restrictions on Anthropic's Claude ... www.thereporter.com
  16. 16 A global workspace in language models www.anthropic.com
  17. 17 Anthropic study reveals what 80000 users want from AI www.edtechinnovationhub.com
  18. 18 Activation Oracles: Training and Evaluating LLMs as General ... alignment.anthropic.com
  19. 19 AI News July 5 2026: Anthropic Enters Drug Discovery ... aitoolsrecap.com
  20. 20 Anthropic shares an approach to agent memory - progress ... www.reddit.com
  21. 21 Anthropic Develops Novel System to Restrict Dual-Use AI Risks blockchain.news
  22. 22 Anthropic Consumer Privacy Policy Takes Effect July 8 techjacksolutions.com
  23. 23 Anthropic Just Dropped Their Claude Code Playbook ... www.youtube.com
  24. 24 Modular Pretraining Enables Access Control openreview.net
  25. 25 Equipping agents for the real world with Agent Skills www.anthropic.com
  26. 26 Mapping AI-enabled cyber threats www.anthropic.com
  27. 27 Anthropic expands partnership with Google and Broadcom ... www.anthropic.com
  28. 28 Anthropic models in Microsoft Online Services learn.microsoft.com
  29. 29 A postmortem of three recent issues www.anthropic.com
  30. 30 Review of the "Risks from automated R&D" section in ... metr.org
  31. 31 AI Security and Access Controls: Best Practices for 2026 ... www.jdsupra.com
  32. 32 Anthropic's Ultracode dynamic workflows for complex tasks www.facebook.com
  33. 33 Promoting Advanced Artificial Intelligence Innovation and ... www.whitehouse.gov
  34. 34 Anthropic has dropped 33-minutes "the prompting ... www.instagram.com
  35. 35 Dual Use Research of Concern research.osu.edu
  36. 36 The Complete Guide to Every Claude Update in Q1 2026 ... aimaker.substack.com
  37. 37 Policies and Safeguards for the Safe Use of AI health-isac.org
  38. 38 A Pretrained Module that Induces an Epistemic Frame via ... arxiv.org
  39. 39 Dual-Use Risks of Frontier Artificial Intelligence papers.ssrn.com
  40. 40 Anthropic AI Attack: How NDR Detects GTG-1002 Cyber ... www.extrahop.com
  41. 41 Data Privacy AI Regulatory and Compliance Update 2026 www.kasowitz.com
  42. 42 Anthropic outage, Claude Code, May 2026: why a 529 ate ... fazm.ai

Schreibe einen Kommentar

Your email address will not be published. Required fields are marked *