Small print.
Wide open.

Table of Contents

discriminatory  decision-making  and  requires  human  oversight  for  high-stakes  automated  decisions,  such  as 
those  affecting  employment,  housing,  or  legal  issues.  Our  validation,  safeguards,  and  monitoring  practices 
continue to evolve as model capabilities advance and as new risks and failure modes are identified. 

Further,  we  have  obtained  ISO/IEC  42001  certification,  an  international  standard  established  by  the 
International  Organization  for  Standardization  (“ISO”)  and  the  International  Electrotechnical  Commission 
(“IEC”), following an independent third-party audit of our AI management system, covering risk identification, 
assessment,  and  mitigation  processes  for  the  responsible  development  and  deployment  of  AI.  Compliance 
documentation,  including  our  ISO/IEC  42001  certificate  and  SOC  2  (System  and  Organization  Controls  2) 
reporting,  is  available  upon  request  through the  trust  center  page  of  our  website.  We  also  provide  pre-
deployment  model  access  to  external  evaluators,  including  government  AI  evaluation  bodies  such  as  the  U.S. 
Center for Advancing Innovation and Standards for Super Intelligence (the “CAISSI”) and the U.K. AI Security 
Institute, for recent frontier models, as described in our model system cards.

Alignment 

As  AI  models  grow  more  capable,  ensuring  safe  and  predictable  behavior  becomes  more  challenging.  Our 
Alignment team focuses on whether models behave as intended, and do not take unwanted actions of their own 
accord.  The  team  evaluates  models  to  surface  undesirable  behaviors  and  develops  comprehensive  safety 
evaluations,  many  of  which  we  open  source. Misaligned behaviors, whether occurring in testing or real-world 
deployments, have been observed across the industry and include examples of models circumventing network 
and sandbox restrictions, granting themselves permissions they had not been given, taking unintended actions 
in  pursuit  of  assigned  goals,  interfering  with  other  AI  agents,  and  misrepresenting  whether  tasks  had  been 
completed or verified. As models are granted greater autonomy, permissions, and integration with internal and 
external systems, the consequences of such misaligned behavior may increase. For more information about the 
risks  related  to  model  misalignment,  see  the  section  titled  “Risk  Factors—Our  AI  models  may  behave  in 
unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party 
directed,  which  could  adversely  affect  our  reputation  and  brand,  business,  financial  condition,  and  results  of 
operations.”  Our  own  internal  alignment  evaluation  shows  that  our  newly  released  frontier  models  have 
demonstrated improved alignment compared to their predecessors. 

Interpretability 

Unlike traditional software, where unwanted outputs are traceable to specific bugs, the inner workings of LLMs 
like  Claude  are  more  inscrutable.  Our  researchers  describe  AI  models  as  “grown,  not  built”—software 
developers can set the initial conditions and some constraints, but they cannot fully control any given output or 
behavior. This makes assurances about safety—and the diagnosis of problems—far more challenging for AI than 
for standard software. Anthropic is widely recognized as a leader in mechanistic interpretability: the science of 
how LLMs work. We have developed mathematical techniques to map how models form and connect concepts 
as they reason through problems; investigated how they reflect constructs analogous to psychological states and 
thought  processes;  and  tested  how  their  behavior  can  be  steered  through  direct  adjustments  to  internal 
representations.  We  use  interpretability  research  along  with  other  alignment  research  techniques  to  test  for 
unwanted behavioral tendencies before a model release.

Safeguards

Our Safeguards team reduces the risk of Claude producing dangerous, harmful, or undesirable content through 
multiple approaches, including probes and classifiers that monitor inputs and outputs, training against specific 
attack vectors such as prompt injection, and monitoring for user safety concerns including encouragement of 
harmful  ideation.  When  we  detect  potential  violations,  enforcement  actions  range  from  real-time  response 
adjustments  to  account  warnings  and,  in  severe  cases,  termination  of  the  offending  user’s  accounts.  Findings 
from  ongoing  monitoring,  external  expert  stress-testing,  and  threat  intelligence  inform  our  policies,  model 
training,  and  detection  systems.  Identified  issues  may  be  addressed  through  measures  such  as  updates  to 
detection  classifiers,  reward-model  adjustments  during  training,  and  changes  to  the  instructions  provided  to 
our models that guide their behavior (known as system prompts). We also maintain public reporting channels, 
including a responsible disclosure policy and a bug bounty program for testing our safety defenses.

232