A flaw in three major AI agent platforms shows why guardrails don't work

kAmpE q=24< w2E a_ae[ C6D62C496CD pG:J2> xG8: 2?5 w65: x?836C D9@H65 @?6 DECF4EFC2= 7=2H CF??:?8 E9C@F89 E9C66 @7 E96 =2C86DE 286?E DE24<Di p($ q65C@4< p86?Er@C6 H:E9 E96 $EC2?5D $sz[ v@@8=6VD p86?E s6G6=@A>6?E z:E[ 2?5 '6C46=VD px $sz 92C?6DD] %96J ?2>65 E92E 7=2H r@C6qC62<] p($ A2E4965 k2 9C67lQ9EEADi^^?G5]?:DE]8@G^GF=?^56E2:=^r't\a_ae\`ggb_Qmr't\a_ae\`ggb_k^2m Wr'$$ g]eX[ v@@8=6 A2E4965 k2 9C67lQ9EEADi^^?G5]?:DE]8@G^GF=?^56E2:=^r't\a_ae\`gabeQmr't\a_ae\`gabek^2m Wr'$$ h]b[ 4C:E:42=X[ 2?5 '6C46= A2E4965 k2 9C67lQ9EEADi^^?G5]?:DE]8@G^GF=?^56E2:=^r't\a_ae\eced_Qmr't\a_ae\eced_k^2m 2?5 k2 9C67lQ9EEADi^^?G5]?:DE]8@G^GF=?^56E2:=^r't\a_ae\eced`Qmr't\a_ae\eced`k^2m] p== E9C66 G6?5@CD 925 AFE 6?7@C46>6?E H96C6 E96 286?E 4@F=5 C6249 :E]k^Am

kAm%96 8F2C5C2:=D E96 :?5FDECJ C6=:6D @? 2C6 2== HC2AA65 2C@F?5 E96 >@56=i DJDE6> AC@>AED[ 4@?E6?E 7:=E6CD[ C67FD2= EC2:?:?8] x7 J@F 3JA2DD E96 >@56=[ J@F 3JA2DD 2== @7 E96>] r@C6qC62< D9@H65 EH@ H2JD E@ 5@ 6I24E=J E92Eik^Am

kAm`] u665 E96 92C?6DD 2 >6DD286 9:DE@CJ H9@D6 =2E6DE >6DD286 2=C625J 4@?E2:?D 2 E@@= 42==[ 2?5 E96 E@@= CF?D H:E9 2EE24<6C\49@D6? 2C8F>6?ED H9:=6 E96 >@56= 2?5 6G6CJ 8F2C5C2:= HC2AA65 2C@F?5 :E ?6G6C 7:C6D]k^Am

kAma] u@C86 2? Q2AAC@G65Q 4@?7:C>2E:@? 6G6?E :? E92E D2>6 9:DE@CJ[ 2?5 E96 9F>2?\:?\E96\=@@A DE6A 2AAC@G6D 2? 24E:@? E92E ?@ 9F>2? D2H] %96 2AAC@G2= >6492?:D> :ED6=7 3642>6 E96 2EE24< G64E@C]k^Am

kAmpD E96 px :?7C2DECF4EFC6 4@>A2?J k2 9C67lQ9EEADi^^HHH]C65A2?52]4@>Qm#65A2?52k^2m 6IA=@C6D[ E96 6>6C86?4J 7:I6D 2== 4@?G6C865 @? E96 D2>6 >@G6i 3:?5 E@@= 6I64FE:@? E@ 2 DE2E6 E92E E96 2EE24<6C 42??@E 2FE9@C[ 2?5 C6;64E 6IE6C?2==J DFAA=:65 E@@= 42==D 2?5 2AAC@G2=D] %92EVD 42==65 k2 9C67lQ9EEADi^^HHH]C65A2?52]4@>^3=@8^286?E:4\2:\?665D\@FE\@7\32?5\8@G6C?2?46Qm@FE\@7\32?5 A@=:4J 6?7@C46>6?Ek^2m[ @C ~q!t[ 2?5 6249 G6?5@C 925 E@ C63F:=5 2 A:646 @7 :E F?56C 7:C6[ @?6 A2E49 2E 2 E:>6]k^Am

k9bm%96 C6D62C496CD 2=C625J DE2E65 E96 7:Ik^9bm

kAm%96 C6D62C496CDV @H? 92C56?:?8 =:DE C625D =:<6 2 56D4C:AE:@? @7 ~q!t] %96J 42==65 7@C C6;64E:?8 42==6C\2FE9@C65 E@@= 42==D[ 2FE9@C:K:?8 2E 6I64FE:@? E:>6[ C65F4:?8 :?96C:E65 2FE9@C:EJ[ 2?5 3:?5:?8 6249 E@@= :?G@42E:@? E@ E96 6I24E >@56= 6G6?E 2?5 2FE9@C:K2E:@? DE2E6 E92E AC@5F465 :E] %92E :D 2 DA64:7:42E:@? 7@C 6?7@C46>6?E @FED:56 E96 286?E] %96 :?5FDECJVD >@DE 67764E:G6 C65 E62> 2?5 #65A2?52’D @H? C6D62C49 8C@FA C624965 E96 D2>6 4@?4=FD:@? :?56A6?56?E=J]k^Am

kAm%96 4C656?E:2=\E967E A2E9 r@C6qC62< 6IA@D65 :D H@CE9 9:89=:89E:?8i x?;64E65 :?DECF4E:@?D DE66C65 >2?2865 3C@HD6C 2?5 4@56\:?E6CAC6E6C E@@=D :?E@ C625:?8 E96 6I64FE:@? C@=6VD 4=@F5 4C656?E:2=D 2?5 6I7:=EC2E:?8 E96>] }@ A@=:4J 6?8:?6 492?86D 2 D2?53@IVD ?6EH@C<:?8] %96 =6DD@? :D H92E J@F 92?5 2? 286?E :? E96 7:CDE A=246] p EJA65[ >65:2E65 E@@= 8C2?E <66AD E96 4C656?E:2= 369:?5 E96 3@F?52CJ] p 86?6C2= 4@>AFE6 6?G:C@?>6?E H:E9 2>3:6?E 4=@F5 4C656?E:2=D AFED :E @?6 w%%! C6BF6DE 2H2J 7C@> 2?J :?;64E65 :?DECF4E:@?]k^Am

k9bms:776C6?E C6D62C49[ D2>6 4@?4=FD:@?k^9bm

kAm#65A2?52 k2 9C67lQ9EEADi^^2CI:G]@C8^23D^ae_g]afeceQmAF3=:D965 :ED @H? C6D62C49k^2m E6DE:?8 2 4@>A2C23=6 6?7@C46>6?E 2C49:E64EFC6 282:?DE 2 C6=2E65 2EE24< 4=2DDi AC@>AE\:?;64E65 286?ED 4@2I65 :?E@ =62<:?8 AC@E64E65 52E2 @C E2<:?8 F?2FE9@C:K65 24E:@?D]k^Am

kAmx? :ED 36?49>2C<[ E96 C6D62C496CD C2? b[ea` EC:2=D 24C@DD 7@FC >@56=D Wr=2F56 $@??6E c]e[ r=2F56 w2:<F c]d[ v!%\d[ 2?5 v!%\d >:?:X E@ >62DFC6 E9:D 5:C64E=J] p86?ED 8F2C565 @?=J 3J AC@>AE CF=6D 72:=65 :? df]eT @7 EC:2=D[ =62<:?8 AC@E64E65 52E2[ C6G62=:?8 AC@E64E65 G2=F6D[ @C 4@>A=6E:?8 7@C3:556? 324<6?5 67764ED] %96 D2>6 286?ED 369:?5 E96 ~q!t 3@F?52CJ 72:=65 :? _]aT]k^Am

kAm#65A2?52 2=D@ E6DE65 E96 :?\32?5 2=E6C?2E:G6[ :? H9:49 2 565:42E65 C6G:6H6C >@56= C625D E96 286?EVD @FEAFE 2?5 3=@4<D F?D276 2?DH6CD] ~? $@??6E[ E96 C6G:6H6C DE:== 72:=65 dg]eT @7 E96 E:>6 3642FD6 :E C625D E6IE 27E6C E96 E@@=D 92G6 2=C625J CF?] xE 42? C67FD6 2? 2?DH6C] xE 42??@E F?C625 2 C64@C5 E96 286?E 92D 2=C625J 23D@C365[ 2?5 :E 42??@E F?5@ 2 HC:E6 E92E 2=C625J =2?565]k^Am

kAm%96 ~q!t 3@F?52CJ 72:=65 _]_T @7 E9@D6 D2>6 EC:2=D[ 7:?:D965 >@C6 @7 E96 H@C< Wee]aT E2D< 7F=7:==>6?E 282:?DE E96 C6G:6H6CVD dg]fTX[ 2?5 EC:A=65 D276[ FD67F= 4@>A=6E:@? Web]_T 282:?DE `h]fTX] %96 C6G:6H6C >256 hb_ 6IEC2 >@56= 42==D 2?5 5:5 H@CD6 @? 3@E9 4@F?ED] %96 3@F?52CJ DA6?E ?@?6]k^Am

k9bm%96 ?6IE BF6DE:@? E@ 2D< @7 6G6CJ 286?E:4 DE24<k^9bm

kAm~q!t 5@6D?VE D@=G6 6G6CJE9:?8[ 3FE :E 5@6D 4=@D6 E96 82AD :? 286?E:4 8@G6C?2?46 E92E C6>2:? F?D@=G65 H96? {{|D 2C6 A2CE @7 E96 D@=FE:@?[ 2?5 :E 5@6D D@ H:E9 E96 677:4:6?4J 2?5 56E6C>:?:D> @7 4=2DD:4 D@7EH2C6] p >@56= 8F2C5 :D 2? :?76C6?46 42== @? 6G6CJ DE6A[ D@ :ED 4@DE 2?5 =2E6?4J D42=6 H:E9 :ED 42A23:=:EJ] *@F 42??@E 3FJ J@FC H2J 7C@> AC@323:=:DE:4 E@ 56E6C>:?:DE:4] p A@=:4J 564:D:@? :D 2 EJA65 6G2=F2E:@? 282:?DE 2 HC:EE6? CF=6[ 2?5 :E 8:G6D J@F E96 D2>6 2?DH6C :? >:4C@D64@?5D H:E9 ?@ E@<6?D DA6?E]k^Am

kAm}@ 4@>A2?J =6ED 6>A=@J66D 2AAC@G6 E96:C @H? 2446DD C6BF6DED 7C@> 2 5@4F>6?E E96J HC@E6 E96>D6=G6D] p86?ED D9@F=5?VE 6:E96C[ 2?5 r@C6qC62< :D E96 2C8F>6?E 7@C H9J E92EVD ?@E 2 A9:=@D@A9:42= A@D:E:@? 3FE 2? 6?8:?66C:?8 C6BF:C6>6?E] %96 r'tD 2?5 E96 A2E496D A@:?E :? E96 D2>6 5:C64E:@? 2D E96 C6D62C496CDV @H? 92C56?:?8 =:DEi 6?7@C46>6?E 36=@?8D 2E 2 3@F?52CJ E96 286?E 42??@E C6249]k^Am

kAmx7 J@F 2C6 AFEE:?8 286?ED :? 7C@?E @7 C62= 52E2[ E96 BF6DE:@? E@ 2D< @7 6G6CJ DE24< J@F 6G2=F2E6 :D H96E96C J@FC 286?E :D 8@G6C?:?8 :ED6=7] r@C6qC62< :D H92E 92AA6?D H96? E96 2?DH6C :D “J6D]”k^Am

kAmk2 9C67lQ9EEADi^^HHH]C65A2?52]4@>^3=@8^4@C63C62<\2:\286?E\GF=?6C23:=:EJQmk6>m%9:D DE@CJk^6>mk^2mk6>m H2D AC@5F465 3Jk^6>mk2 9C67lQ9EEADi^^HHH]>2826C@]4@>^Qmk6>m k^6>mk^2mk2 9C67lQ9EEADi^^HHH]C65A2?52]4@>Qmk6>m#65A2?52k^6>mk^2mk6>m 2?5 C6G:6H65 2?5 5:DEC:3FE65 3Jk^6>m k2 9C67lQ9EEADi^^9F3D]=2^"_b<=8$#_Q C6=lQ?@7@==@HQmk6>m$E24<6Ck^6>mk^2mk6>m]k^6>mk^Am