,,ii #aaith ,,congress #b;d ,session ,s4 #ebeb ,,in ,,the ,,senate ,,of ,,the ,,united ,,states ,august #e1 #bjbf,mr4 ,hagerty 7for himself1 ,mr4 ,kim1 ,mr4 ,scott of ,south ,carolina1 and ,ms4 ,cortez ,masto7 introduced the following bill2 which was read twice and referred to the ,committee on ,banking1 ,housing1 and ,urban ,affairs ,a ,,bill ,to prevent foreign adversaries from threatening the national security of the ,united ,states by extracting key technical features of closed-source1 ,united ,states-owned artificial intelligence models1 and for other purposes4 #a4 ,short title ,this ,act may be cited as ,blocking ,large-scale ,adversarial ,distillation ,efforts ,act of #bjbf or ,,blade ,act4 #b4 ,sense of ,congress ,it is the sense of ,congress that-- 7#a7 artificial intelligence models owned by ,united ,states private sector entities are essential for advancing economic and national security interests of the ,united ,states2 7#b7 many of the most advanced artificial intelligence models owned by ,united ,states entities are closed-source models whose unique technical characteristics are not openly shared or published2 7#c7 the unauthorized acquisition of model capabilities1 such as model weights1 model architectures1 and other technical characteristics of closed-source artificial intelligence models1 by persons of concern through model extraction attacks represents a threat to the national security and foreign policy interests of the ,united ,states1 as well as the intellectual property rights and economic competitiveness of ,united ,states entities2 7#d7 the ,united ,states ,government1 in cooperation with private owners of closed-source artificial intelligence models1 should take steps to identify1 punish1 and deter model extraction attacks on the protected capabilities of closed-source artificial intelligence models by persons of concern2 7#e7 model extraction attacks against ,united ,states closed-source artificial intelligence models allow foreign adversaries a short cut to acquiring advanced artificial intelligence capabilities2 and 7#f7 authorized model training practices that adhere to the terms of service or are otherwise consistent with contractual terms set by the owners of closed-source artificial intelligence models are a legitimate research method that play an important role in artificial intelligence research and are fundamentally distinct from model extraction attacks addressed by this ,act4 #c4 ,definitions ,in this ,act3 7#a7 ,appropriate congressional committees ,the term appropriate congressional committees means-- 7,a7 the ,committee on ,foreign ,affairs of the ,house of ,representatives2 and 7,b7 the ,committee on ,banking1 ,housing1 and ,urban ,affairs of the ,senate4 7#b7 ,closed-source artificial intelligence model ,the term closed-source artificial intelligence model means any artificial intelligence model with the following characteristics3 7,a7 ,proprietary key technical information1 such as underlying model weights1 that are necessary to reproduce and independently recreate the model and that are not willingly shared with third parties or otherwise made publicly available by the owner of the model4 7,b7 ,access and use governed by terms of service or contractual agreements that are established by the owner of the model4 7,c7 ,access that is provided via an application program interface or another consumer-facing1 owner-controlled interface without enabling third parties to obtain1 modify1 or host the closed-source artificial intelligence model on their own data servers or other technology unless specifically authorized by the owner of the model4 7#c7 ,country of concern ,the term country of concern means-- 7,a7 the ,people's ,republic of ,china1 including the ,hong ,kong and ,macau ,special ,administrative ,regions2 7,b7 the ,russian ,federation2 and 7,c7 any other foreign country-- 7i7 listed in ,country ,group ,d3#e in ,supplement ,no4 #a to part #gdj of title #ae1 ,code of ,federal ,regulations1 as published on ,january #a1 #bjbf1 that is designated by the ,secretary of ,commerce as a country of concern for purposes of this section and for which notice of such designation has been published in the ,federal ,register2 and 7ii7 identified by the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security1 pursuant to an assessment required by subsection 7a7 or 7e7 of section #d4 7#d7 ,person of concern ,the term person of concern means any foreign person that-- 7,a7 is located or headquartered in1 or the ultimate parent company of which is headquartered in1 a country of concern2 7,b7 is operating under the direction or control of any entity located or headquartered in1 or the ultimate parent company of which is headquartered in1 a country of concern2 or 7,c7 is conducting or attempting to conduct a model extraction attack against closed-source artificial intelligence models owned by ,united ,states persons and outside of authorized model training practices4 7#e7 ,foreign person ,the term foreign person means a person that is not a ,united ,states person4 7#f7 ,fraudulent account network provider 7,a7 ,in general ,the term fraudulent account network provider means any foreign person that knowingly and intentionally creates1 obtains1 maintains1 sells1 brokers1 or otherwise provides access to an account that allows a person of concern to access a closed-source artificial intelligence model that the entity would otherwise be prohibited from accessing as a result of location restrictions in the terms of service or a contractual agreement created by the owner of the model4 7,b7 ,exception ,for purposes of subparagraph 7,a71 an entity that creates or transmits location information to enable persons within countries of concern to access the internet for purposes of freedom of expression is not1 on the basis of that activity alone1 a fraudulent account network provider4 7#g7 ,model extraction attack 7,a7 ,in general ,the term model extraction attack means the unauthorized extracting of the capabilities of a closed-source artificial intelligence model to replicate1 develop1 train1 or improve another artificial intelligence model1 if such extraction-- 7i7 circumvents technical1 contractual1 or other access controls1 identity verification requirements1 or geographic access restrictions implemented by the owner of the model2 7ii7 is conducted through fraudulent1 misrepresented1 or unauthorized credentials2 or 7iii7 violates the terms1 conditions1 or restrictions governing access to or use of the model1 as established by the owner1 that specifically prohibit the use of model outputs or interactions to replicate1 develop1 train1 or improve another artificial intelligence model4 7,b7 ,inference of purpose ,for purposes of subparagraph 7,a71 the purpose of extraction may be inferred from the totality of circumstances1 including-- 7i7 the volume1 structure1 pattern1 coordination1 or timing of the extraction activity2 7ii7 the concentration of extractions on specific model capabilities2 7iii7 the use of multiple accounts in a coordinated manner2 or 7iv7 the correlation of extraction activity within the development timeline of another artificial intelligence model4 7,c7 ,exclusion ,for purposes of subparagraph 7,a71 model training activities conducted in compliance with the terms1 conditions1 and restrictions governing access to and use of a closed-source artificial intelligence model1 or otherwise conducted within a permitted exception or the express authorization of the owner of the model1 are not model extraction attacks4 7#h7 ,operating ,committee for ,export ,policy ,the term ,operating ,committee for ,export ,policy means the ,operating ,committee for ,export ,policy referred to in section #agfc7;c7 of the ,export ,control ,reform ,act of #bjah 7#ej ,u4,s4,c4 #dhbb7;c774 7#i7 ,owner ,the term owner means1 with respect to a closed-source artificial intelligence model1 the person that-- 7,a7 holds intellectual property rights 7including trade secret1 copyright1 patent1 or other proprietary rights71 contractual rights1 or a combination thereof1 sufficient to authorize or restrict third-party access to1 use of1 extraction from1 or reproduction of the model1 or any version1 instance1 or deployment the model1 whether such rights were obtained through development1 acquisition1 assignment1 license1 or otherwise2 and 7,b7 is a ,united ,states person4 7#aj7 ,person ,the term person means individual or entity4 7#aa7 ,united states person ,the term ,united ,states person means-- 7,a7 a ,united ,states citizen or an alien lawfully admitted for permanent residence to the ,united ,states2 7,b7 an entity organized under the laws of the ,united ,states or any jurisdiction within the ,united ,states1 including a foreign branch of such an entity2 or 7,c7 any person located in the ,united ,states4 #d4 ,assessment of model extraction attacks and fraudulent account network providers 7a7 ,in general ,not later than #ahj days after the date of the enactment of this ,act1 the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall complete an assessment to determine-- 7#a7 which1 if any1 persons of concern have conducted or are currently conducting model extraction attacks against closed-source artificial intelligence models owned by ,united ,states persons2 and 7#b7 which1 if any1 persons of concern are fraudulent account network providers4 7b7 ,matters ,to be included ,the assessment required by subsection 7a7 shall include the following3 7#a7 ,a determination of which persons of concern-- 7,a7 have either previously or are currently engaging in model extraction attacks2 or 7,b7 are fraudulent account network providers4 7#b7 ,a determination of the countries1 if any-- 7,a7 from which model extraction attacks have originated2 and 7,b7 in which fraudulent account network providers exist4 7#c7 ,an identification of which1 if any1 agencies or instrumentalities of governments of countries of concern have provided or are providing material assistance to entities identified pursuant to paragraph 7#a74 7#d7 ,an analysis of the methods employed by persons of concern identified pursuant to paragraph 7#a71 including-- 7,a7 the role of fraudulent account network providers in model extraction attacks1 including1 to the extent possible1 the physical location of offices and data centers of such providers2 and 7,b7 a determination1 to the extent possible1 of the number of attempted model extraction attacks that occurred during the #b calendar years preceding the date on which the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security1 begins the assessment required by subsection 7a74 7#e7 ,an examination of the strengths and weaknesses of various detection approaches that can be used to determine whether a model extraction attack has occurred or is occurring4 7#f7 ,an assessment of the economic and national security consequences of successful model extraction attacks by persons of concern that occurred during the #b calendar years preceding the date on which the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security1 begins the assessment required by subsection 7a74 7#g7 ,steps detailing how the ,united ,states ,government is assisting owners of closed-source artificial intelligence models that have been the target or victim of model extraction attacks in detecting model extraction attacks1 deterring future model extraction attacks1 and punishing persons of concern that engage in model extraction attacks or are fraudulent account network providers4 7#h7 ,a diplomatic strategy to leverage allies and partners of the ,united ,states in detecting and preventing model extraction attacks by persons of concern4 7c7 ,public consultation 7#a7 ,in general ,in conducting the assessment required by subsection 7a71 the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall consult with owners of closed-source artificial intelligence models that have been the targets or victims of model extraction attacks1 academic experts1 industry fora1 and other appropriate entities-- 7,a7 to identify patterns of behavior and methods of attackers to better inform efforts of the ,united ,states ,government and the private sector to detect model extraction attacks2 7,b7 to develop best practices for defending against model extraction attacks2 and 7,c7 to develop best practices for identifying activities of fraudulent account network providers that facilitate model extraction attacks4 7#b7 ,voluntary participation ,the participation of owners of closed-source artificial intelligence models described in paragraph 7#a7 in consultations under that paragraph shall be voluntary4 7d7 ,report 7#a7 ,in general ,not later than #baj days after the date of the enactment of this ,act1 the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall submit to the appropriate congressional committees a report that contains the findings of the assessment required by subsection 7a74 7#b7 ,updates ,not later than one year after submitting the report required by paragraph 7#a71 and annually thereafter for #c years1 the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security1 shall submit to the appropriate congressional committees an update to the report listing any additional persons of concern identified pursuant to subsection 7a74 7#c7 ,form ,the report required by paragraph 7#a71 and each update required by paragraph 7#b71 shall be submitted in unclassified form1 but may contain a classified annex4 7e7 ,routine assessment ,the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall routinely assess for-- 7#a7 model extraction attacks directed against owners of closed-source artificial intelligence models that occur after the date of completion of the assessment required by subsection 7a72 7#b7 fraudulent account network providers that facilitate model extraction attacks after that date2 and 7#c7 any material changes related to other matters specified in subsection 7b74 7f7 ,industry coordination ,the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall establish an information-sharing mechanism that allows owners of closed-source artificial intelligence models to voluntarily1 quickly1 and confidentially share information about model extraction attacks and fraudulent account network providers with the ,department of ,commerce4 7g7 ,,ai model extraction attackers list 7#a7 ,in general ,the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall-- 7,a7 maintain a list1 to be known as the ,,ai ,model ,extraction ,attackers ,list1 that displays information about specific persons of concern identified pursuant to an assessment required by subsection 7a7 or 7e7 as having conducted or directed model extraction attacks in the past year2 and 7,b7 publish the list on a publicly available website of the ,department of ,commerce4 7#b7 ,protection of confidential information ,the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security1 may not1 in publishing the list required by paragraph 7#a7 on a publicly available website of the ,department of ,commerce1 disclose confidential information provided by the owner of a closed-source artificial intelligence model without the express permission of the owner4 7h7 ,public guidance 7#a7 ,in general ,not later than #baj days after the date of the enactment of this ,act1 the ,secretary of ,commerce1 acting through the ,under ,secretary of ,commerce for ,industry and ,security and in coordination with the head of each agency that is a member of the ,operating ,committee for ,export ,policy1 shall publish a report comprising of best practices to detect1 prevent1 and respond to model extraction attacks4 7#b7 ,public access ,the report required by paragraph 7#a7 shall be publicly available4 7#c7 ,protection of confidential information ,in making the report required by paragraph 7#a7 publicly available1 the ,secretary may not disclose confidential information provided by the owner of a closed-source artificial intelligence model without the express permission of the owner4 #e4 ,deterring model extraction attacks and fraudulent account network providers 7a7 ,consideration of addition to ,entity ,list 7#a7 ,in general ,not later than #baj days after the date of the enactment of this ,act1 the ,under ,secretary of ,commerce for ,industry and ,security1 in coordination with the head of each agency that is a member of the ,end-,user ,review ,committee1 shall add each entity described in paragraph 7#b7 to the ,entity ,list maintained by the ,bureau of ,industry and ,security and set forth in ,supplement ,no4 #d to part #gdd of title #ae1 ,code of ,federal ,regulations1 or any successor regulations4 7#b7 ,entities described ,an entity described in this paragraph is-- 7,a7 an entity that is a person of concern identified1 under subsection 7a7 or 7e7 of section #d1 as having conducted model extraction attacks or having facilitated such attacks via fraudulent account networks2 or 7,b7 a subsidiary of such an entity 7to be determined by ownership of #ej percent or more in the aggregate1 directly or indirectly74 7b7 ,imposition of sanctions 7#a7 ,in general ,the ,president shall1 pursuant to the ,international ,emergency ,economic ,powers ,act 7#ej ,u4,s4,c4 #agja et seq471 block and prohibit all transactions in all property and interests in property of each person of concern identified under subsections 7a7 and 7e7 of section #d if such property and interests in property are in the ,united ,states1 come within the ,united ,states1 or are or come within the possession or control of a ,united ,states person4 7#b7 ,exceptions 7,a7 ,exception relating to the provision of humanitarian assistance ,sanctions under paragraph 7#a7 may not be imposed with respect to transactions or the facilitation of transactions for-- 7i7 the sale of agricultural commodities1 food1 medicine1 or medical devices2 7ii7 the provision of humanitarian assistance2 or 7iii7 transporting goods or services that are necessary to carry out operations relating to humanitarian assistance4 7,b7 ,exception for intelligence1 law enforcement1 and national security activities ,sanctions under this subsection shall not apply to any authorized intelligence1 law enforcement1 or national security activities of the ,united ,states4 7#c7 ,waiver 7,a7 ,in general ,the ,president may1 subject to subparagraph 7,b71 waive the application of subsection 7a7 or 7b7 with respect to a person if1 before issuing the waiver1 the ,president submits to ,congress-- 7i7 a certification in writing that the issuance of the waiver is in the national interests of the ,united ,states2 and 7ii7 a report explaining the basis for the certification4 7,b7 ,form of report ,each report required by subparagraph 7,a7 shall be submitted in unclassified form but may include a classified annex4 7#d7 ,penalties ,a person that violates1 attempts to violate1 conspires to violate1 or causes a violation of this subsection or any regulation1 license1 or order issued to carry out that subsection shall be subject to the penalties set forth in subsections 7b7 and 7c7 of section #bjf of the ,international ,emergency ,economic ,powers ,act 7#ej ,u4,s4,c4 #agje7 to the same extent as a person that commits an unlawful act described in subsection 7a7 of that section4