Powe ful ope models ca be t ai ed fo a a ge of safety a d secu ity challe ges without comp omisi g capability. I flectio AI a d AI21 post-t ai ed NVIDIA Nemot o ope models with Alice’s safety data, demo st ati g a path to espo sibly deployi g customizable AI.
NEW YORK, Oct. 15, 2026 (GLOBE NEWSWIRE) — Alice a ou ced the esults of two collabo atio s focused o alig i g NVIDIA Nemot o models to compa y-specific safety a d secu ity policies. The esults demo st ate how eal-wo ld th eat data ca be used to t ai ope -sou ce models fo f o tie safety pe fo ma ce.
Ope -sou ce models give o ga izatio s g eate co t ol ove AI behavio , allowi g them to t ai policies i to the u de lyi g model athe tha setti g i st uctio s as system p ompts. To successfully t ai models o thei policies, compa ies eed adve sa ial data a d expe tise. Two p ojects demo st ate how NVIDIA’s state-of-the-a t ope -sou ce model ca be alig ed fo safe huma i te actio s, a d mo e secu e e te p ise age ts.
“Powe ful AI is bei g democ atized. It’s ow available to compa ies eve ywhe e, a d they wa t to use it safely a d p edictably,” said Noam Schwa tz, CEO a d co-fou de of Alice. “These esults show that with the ight eal-wo ld adve sa ial data a d t ai i g, we ca teach ope sou ce models how to ha dle the th eats a d abuse they will actually face, ac oss ve y diffe e t safety challe ges, while keepi g them as powe ful. That’s the ba ow: f o tie -g ade safety that a y o ga izatio u i g adva ced AI ca actually put i to p actice.”
I flectio AI is buildi g pe so al AI that empowe s people a d shifts i te actio s f om t a sactio al to elatio al. Allowi g mo e atu al i te actio s with co ve satio al AI also i c eases isks which most be chma ks ca ’t catch. I the p ocess of buildi g the ext ge e atio of pe so al i tellige ce, I flectio AI fi e-tu ed NVIDIA Nemot o 3 Supe 120B-A12B, a d tu ed to Alice to develop ed-teami g datasets aimed at educi g sycopha cy a d p eve ti g the model f om ei fo ci g u healthy attachme t ove lo g co ve satio s. Fi e-tu i g o that data aised Pi’s alig me t sco e by 30% while keepi g u ecessa y efusals to just 1.6%.
Alice c eated mo e tha 20 ealistic pe so as a d used them i multi-tu ed-team exe cises whe e those isks could develop ove time. The pe so as i cluded vul e able but be ig use s as well as adve sa ial use s delibe ately attempti g to ma ipulate the model, d awi g o Alice’s eal-wo ld k owledge of ha mful a d adve sa ial behavio .
“As AI i ovatio accele ates, safety has to adva ce apidly as well. It’s o lo ge e ough to ask whethe a model ca ide tify a d efuse a ha mful p ompt,” said Sea White, CEO of I flectio AI. “We also eed to u de sta d how AI behaves ac oss hu d eds of i te actio s, whethe it challe ges athe tha simply ag ees, if it suppo ts huma age cy athe tha depe de cy, a d ultimately if it leaves people mo e capable. This wo k with Alice a d NVIDIA demo st ates that we ca t ai those p i ciples i to powe ful ope models without sac ifici g what makes them useful. That’s fu dame tal to the ki d of huma -ce te ed Pe so al I tellige ce we’ e buildi g with Pi.”
Alice also wo ked with AI21, which has t ai ed la guage models si ce 2021 a d ow offe s lab-as-a se vice fo specialized i tellige ce. The wo k with them focused o a ve y diffe e t p oblem: p otecti g e te p ise age ts agai st i di ect p ompt i jectio . As age ts gai access to compa y systems, data a d tools, attacke s ca attempt to ma ipulate them th ough i fo matio the age t e cou te s while doi g its wo k. T ai i g agai st those attacks the efo e ca ot focus o ly o whethe the model efuses a malicious i st uctio . A secu e age t also eeds to ecog ize the attack while co ti ui g to complete the legitimate task.
Alice’s ei fo ceme t-lea i g e vi o me ts we e built to t ai a d measu e both. AI21 used them to post-t ai the ece tly eleased NVIDIA Nemot o 3.5 Supe VL ope model, i c easi g the model’s esilie ce to i di ect p ompt i jectio by 41 poi ts, competi g with GPT-6.1 Sol.
The esult shows that secu ity a d capability a e ot ecessa ily competi g objectives. With the ight t ai i g, models ca become sig ifica tly mo e esilie t to adve sa ial ma ipulatio while mai tai i g task pe fo ma ce.
“Fo most compa ies, the ext phase of AI adoptio goes beyo d f o tie models, to ope models they ow , shape a d co t ol,” said O i Goshe , Co-fou de a d CEO of AI21. “Ow i g a model mea s ow i g how it fails, ot o ly what it ca do. We post-t ai ed NVIDIA Nemot o 3.5 Supe -VL o attack t aces Alice sees i the wild, so it holds up agai st the attacks that actually each e te p ises while keepi g the behavio it was built fo .”
“NVIDIA Nemot o p ovides the ope weights, ecipes a d tooli g fo o ga izatio s to post-t ai o thei ow data a d optimize accu acy a d pe fo ma ce fo thei specific wo kloads,” said Joey Co way, se io di ecto Nemot o a d Softwa e, NVIDIA. “Alice’s wo k with I flectio AI a d AI21 shows how post-t ai i g Nemot o o eal-wo ld adve sa ial data delive s powe ful safety a d secu ity — exactly the ki d of i ovatio ope models a e mea t to u lock.”
Take togethe , these collabo atio s demo st ate how eal-wo ld adve sa ial i tellige ce, safety datasets, expe t ed teami g a d ei fo ceme t-lea i g e vi o me ts ca tu ope models i to systems t ai ed fo the isks they will actually face. The b oade oppo tu ity is clea : powe ful, customizable AI that compa ies ca ow a d deploy with f o tie -level safety built i .
Alice is a AI safety a d secu ity compa y helpi g model builde s a d e te p ises t ai , evaluate, a d deploy safe AI. Its adve sa ial i tellige ce combi es co ti uously eme gi g th eats f om ope a d closed commu ities with expe t esea ch to powe ed teami g, evaluatio s, t ai i g data, RL e vi o me ts, a d e te p ise gua d ails. Alice wo ks with 8 of the 10 leadi g fou datio model labs a d AI- ative compa ies, while helpi g e te p ises test a d p otect AI applicatio s agai st eal-wo ld safety a d secu ity isk.
AI21 is a AI lab buildi g f o tie models a d AI systems fo the e te p ise si ce 2017. C eato of the Ju assic a d Jamba model families, AI21 today offe s its lab-as-a-se vice fo specialized i tellige ce: Fo wa d-Deployed Resea che s a d a ha ess- a d model-t ai i g platfo m that tu ope fou datio s i to models its custome s ow , delive ed with the weights, ecipes a d evaluatio s that keep them thei s. AI21 has secu ed $336 millio i fu di g f om i dust y leade s, i cludi g Google, a d I tel, ei fo ci g its commitme t to adva ci g AI i ovatio .
CONTACT: Media Co tact:
Ma k P i dle
ma k.p i dle@fusio p .com
