Fou datio models such as Ope AI’s GPT, A th opic’s Claude, a d Google’s Gemi i have adva ced at a ext ao di a y pace ove the past few yea s, pa ticula ly i codi g a d softwa e easo i g. These systems ca ow w ite a d efi e softwa e, ide tify a d emediate bugs, easo th ough complex e vi o me ts, a d o chest ate tools ac oss multi-step wo kflows. Those same capabilities map di ectly to offe sive secu ity. Vul e ability esea ch a d exploit developme t a e, at thei co e, code- easo i g p oblems executed agai st bou ded ta gets. Whe combi ed with age tic executio , these models move beyo d assisti g attacke s to auto omously ca yi g out e d-to-e d attack chai s, as demo st ated i A th opic’s Mythos disclosu e.
But defe sive cybe secu ity – eve ythi g outside of code secu ity – dema ds a fu dame tally diffe e t set of capabilities. Rathe tha code easo i g a d ge e atio , it equi es diggi g th ough massive volumes of secu ity data (such as audit logs, eve ts, a d etwo k flows), co elati g weak sig als ove lo g time ho izo s, a d mai tai i g ext eme co siste cy ac oss thousa ds of decisio s i seque ce.
Co ma’s esea ch shows just that. Usi g ealistic e te p ise e vi o me ts modeled afte Fo tu e 500 o ga izatio s, complete with the doze s of secu ity tools typically deployed i la ge e te p ises, Co ma a hu d eds of simulatio s with leadi g AI models such as Ope AI’s GPT a d A th opic’s Claude. Fi st, the models we e tasked with acti g as attacke s, pla ti g pe siste t th eats i side e te p ise systems. The same models we e the asked to defe d those e vi o me ts, locati g a d emediati g the th eats they had c eated. The esults we e co siste t: i ea ly eve y case, the same AI model that executed a e d-to-e d attack could ‘t defe d agai st the ve y attack it ca ied out – AI attacke s succeeded 88% of the time, while AI defe de s detected just 12%. This highlights a g owi g imbala ce with majo implicatio s fo cybe secu ity.
“The ace to ge e al i tellige ce i cybe secu ity has al eady begu , a d the attacke s have a sig ifica t head sta t,”
Co ma is buildi g the fi st fou datio model pu pose-built fo defe sive cybe secu ity. It powe s Co ma’s AI age ts, allowi g them to outpe fo m age ts built o othe models while ge e alizi g ac oss the full spect um of defe sive secu ity tasks. O ga izatio s o boa d Co ma much like they would a ew team membe . O ce deployed, Co ma’s age ts ca ope ate ac oss vi tually eve y a ea of defe sive cybe secu ity, co ti uously lea the e vi o me t a ou d them, a d scale to meet dema ds that o team could cove alo e.
Si ce lau chi g just six weeks ago, Co ma’s AI wo kfo ce has bee deployed at Fo tu e 100 a d Fo tu e 500 o ga izatio s ac oss healthca e, fi a cial se vices, e e gy, c itical i f ast uctu e, etail, a d othe secto s. Ea ly deployme ts have educed th eat espo se times by mo e tha 94%, expa ded secu ity cove age by 15 times ac oss diffe e t secu ity fu ctio s, a d u cove ed multi-stage attack campaig s that would have othe wise go e u detected.
“Co ma has t ai ed its model fo the complexity of eal-wo ld attacks a d is buildi g the i tellige ce laye defe se actually eeds,”
“AI is eshapi g cybe secu ity i ways that i c easi gly exte d beyo d the e te p ise to atio al secu ity a d geopolitical stability,”
Co ma b i gs togethe a fi st-of-its-ki d combi atio of expe tise ac oss p e-t ai i g, post-t ai i g, offe sive a d defe sive cybe secu ity – u iti g f o tie AI esea che s f om Google a d DeepMi d with cybe secu ity expe ts f om the most elite g oups withi Is ael’s 8200 U it a d the wo ld’s leadi g cybe secu ity compa ies. It is the type of i te discipli a y team equi ed to tackle Co ma’s missio , lo g co side ed “the holy g ail of cybe secu ity.”
Co tact i fo:Yuval Po atMedia Co sulta t
