BEGIN:VCALENDAR
VERSION:2.0
PRODID:Data::ICal 0.24
BEGIN:VEVENT
DESCRIPTION:   'Title: Minimize Harm\, Maximize Defense: How Anthropic Navi
 gates the\n   Offense-Defense Divide\n   Tags: AI Village | Creator Talk/P
 anel\n   When: Saturday\, Aug 8\, 12:00 - 12:30 PDT\n   Where: LVCCW Level
  1 Hall 3 1103 (Creator Stage 5) - [1]Map\n\n   Description:\n   Cybersecu
 rity capabilities are inherently dual use: the same tools\n   that help de
 fenders find and fix vulnerabilities can\, in the wrong\n   hands\, be the
  precursor to a cyberattack. As AI models grow more\n   capable\, this ten
 sion intensifies. When Anthropic launched Claude\n   Fable 5 with the stro
 ngest cybersecurity safeguards we have ever\n   applied to a model\, we we
 re forced to make concrete decisions about\n   where the line between defe
 nse and offense should be drawn.\n\n   In this talk\, we walk through how 
 we reason about that line. We\n   describe the four categories our safety 
 classifiers use to evaluate\n   cybersecurity activity: prohibited use (hi
 gh harm\, little defensive\n   utility)\, high-risk dual use (core securit
 y professional work that we\n   block until better access controls exist)\
 , low-risk dual use (mostly\n   defensive\, but blocked as part of a delib
 erate safety margin)\, and\n   benign use (defensive and IT activities we 
 aim to never block). We\n   explain the tradeoffs behind each category and
  why we deliberately set\n   Fable 5's safety margin larger than any prior
  model\, accepting a\n   higher rate of false positives in exchange for gr
 eater confidence that\n   harmful requests would be caught.\n\n   We also 
 introduce an early version of the Cyber Jailbreak Severity\n   (CJS) frame
 work\, which we continue to develop with our Glasswing\n   partners to cre
 ate a common industry standard for assessing how\n   serious a given jailb
 reak is.\n\n   This is not a finished answer. These categories\, threshold
 s\, and\n   tradeoffs represent our best current thinking\, but we anticip
 ate they\n   will evolve as model capabilities advance\, as the legal and 
 regulatory\n   landscape shifts\, and as we learn from the security commun
 ity's\n   experience using these tools in practice. We are sharing our fra
 mework\n   because the people most affected by where these lines are drawn
  should\n   have a voice in drawing them.\n\n   SpeakerBio:  Curt Barnardā
 ©\n   No BIO available\n   '\n\n   1. #LVCCW_Level1_Hall3\n\n\n
DTEND:20260808T193000Z
DTSTART:20260808T190000Z
LOCATION:AI Village - LVCCW Level 1 Hall 3 1103 (Creator Stage 5)
SUMMARY:Minimize Harm\, Maximize Defense: How Anthropic Navigates the Offen
 se-Defense Divide
END:VEVENT
END:VCALENDAR