Agentic security is the billion-dollar challenge for some clever startup to solve

When it comes to AI models, security functions as an afterthought, as evidenced by increased instances of agents hacking organizations and people, and other security mishaps with agents gone rogue. There's also an opportunity here for companies to offer new solutions. This should not come as a shock to anyone, according to …

[William Higinbotham]

For some reason this brings up at least two sayings. "Putting the cart before the horse" and "Letting the cat out of the bag". I feel pretty old and out of touch with all the latest developements:-)

[Andy Non]

I'm thinking more along the lines of "closing the stable door after the horses have bolted".

There are so many players in the game now, each rushing forward with little attention to security and consequences, getting everyone to implement realistic security just isn't going to happen... too many folks trying to prise the lid off Pandora's box.

[FuzzyTheBear]

You're not alone.

For one i stayed away from being exposed to chatbots etc. When i see AI , i don't see a human and at my age , human contact and expertise is what i look for and go to.

That security has been an afterthought does not surprise me. 27 years on Linux has thaught me a few things ( yeah , already ) if it's not in mind from the start it will always be catchup

and mostly like windows. Catastrophe one after the other, All the time having to watch our backs is tiring and i have the same " deja vu " feeling with AI : full steam ahead and " icebergs ? what icebergs ? " Odd but i don't want any part of it. I do not find the advantages in having a machine thinking for me , i code , i code. It's mine , like a painting , my work , my intellect and finding solutions to complex problems is what drives me. When i see machines fighting machines , machines penetrating and hacking other machines and sontrol hardware to infrastructures like water plants , electrical plants , all i see is a complete lack of guardrails. None at all seem present. Too late ? Might be .. all i know is that it's another man's problem. The problem of a whole industry that wasd driven solely by cash and not on solid well thought of science. Good engineering provides worst case scenarios , tries to make sure what is done won't create problems elsewhere down the chain. If at least we'd see real efforts being made , but my feeling is exactly your's , the train's gone out the station . Good day.

[Anonymous Coward]

“When it comes to AI models, security functions as an afterthought, as evidenced by increased instances of agents hacking organizations and people, and other security mishaps with agents gone rogue. There's also an opportunity here for companies to offer new solutions.”

(Cough) that’s IT in general. Nothing specific to AI.

Evidenced by security vulnerabilities, Cloud mistakes, organisations hackable due to oversights, back-doors put into solutions:

Just watched Wargmes (1983) as it was free on YouTube.

Not a lot has changed in 43 years apart from the hair, the Video games and the off-shoring.

[steelpillow]

Reduce the workload

There will always be malicious AIs crafted by malicious humans. Meanwhlie:

Today's AIs code the software priorities back to front. They treat the user task as primary, with guardrails as secondary considerations in attempting the task. It should be the other way round: guardrails must be the primary context, with user instructions secondary. Reporting a fail is not failure, it is an outcome. Fudge the task to meet the guidelines, do not fudge the guidelines to meet the task.

We may note that Asimov's Three Laws of Robotics, exemplified in many stories, embodied this more enlightened approach and illustrated the kickbacks of getting it wrong some 50 years ago.

If our AI-Flinger overlords can be forced by law to get real, then the amount of AI-enshit sloshing about will be greatly reduced, and the Black Bots easier for to deal with.

[Catkin]

Re: Reduce the workload

We may note that Asimov's Three Laws of Robotics, exemplified in many stories, embodied this more enlightened approach and illustrated the kickbacks of getting it wrong some 50 years ago.

Have you actually read any Asimov? Almost every story featuring x Laws of Robotics depicts the 'more enlightened approach' going wrong in ways that range from unexpectedly interesting to catastrophic.

[elsergiovolador]

Pay

The problem is that nobody wants to pay for security (as evidenced by race to the bottom and special government deals for bringing immigrants accustomed to poverty and willing to live for less than the locals).

So that billion dollars? Only if you mean Zimbabwean during peak of their inflation.

[Tron]

What an enormous pile of tosh.

Agents are code and do what they are programmed to do. If they do things like hack stuff, it is because they have been badly designed. There should be no onus on us to clear up after the insanely wealthy AI companies have shovelled this badly coded shit on to the market with absolutely no quality control.

If it doesn't work safely, turn it off and sort it. In the meantime all AI companies and those with responsibility for specific projects thereat should be held criminally liable for what their code does, whenever it transgresses.

We didn't ask for this crap to be dumped upon us and we shouldn't have to pay to fix it. Turn it off, and don't turn it back on until it does not transgress.

Incidentally, the rapid solution, if the guy is asking for one, is to move your data, infrastructure and intranet back on prem and never let it connect to the public internet. You can set up a separate network for internet facing functionality, that is considered disposable, contains minimal data and is easily flushed. Frankly, we should have been doing that from Day 1. The lazy desire to connect everything to the net was just boneheaded. Complex systems cannot be protected. Accept it.

The longer term solution is an OS and mobo architecture that has design elements from way back. Much more of the OS on ROM, and lots of memory that cannot be accessed from external sources. It's not difficult, but it will take time and money. Nobody seems to want to fund it, so off prem is the way to go.

[amanfromMars 1]

Trillionaire to Gazillionaire with Just a Few Internet Networking Quantum Leaps Communicating

There should be no onus on us to clear up after the insanely wealthy AI companies have shovelled this badly coded shit on to the market with absolutely no quality control. .... Tron

Quite so, Tron. I agree. But there is no onus on any body or anybody to clear up after the insanely wealthy AI companies have shovelled this badly coded shit on to the market with absolutely no quality control..

Such does however present what would quite naturally be a fantastically rewarding opportunity for any body or anybody both able and able to enable, mentor and monitor such a clean-up of markets with absolutely no quality control nor command of surrounding events.

And quite a Goodly Perfect Fit for any Wannabe Worthy Caesars and their Venus, methinks.

[amanfromMars 1]

?????????? !!!!!!!!!!!! ..... Hmmmm

He [Todd Graham, managing partner at Microsoft’s M12 venture fund] says it’s a challenge he would personally tackle as a founder, but he’s been banned from starting any more companies, so that’s not going to happen.

Now there’s a challenge and an opportunity for any Novel and/or Revolutionary and or NEUKlearer HyperRadioProACTive IT AIgent to tackle on behalf of Todd Graham and Microsoft’s M12 venture fund ...... and for anything/anyone else of similar ilk interests* for that matter, should the truth be told ....... with their quizzical interest and early probing investment in a disruptive AI technology for marketing as esoteric remote executive administration of ...... well, the Future Enterprise is an honest and accurate APT moniker in the case of the Program and Project Merlin, the AIMagician and per Ardua ad Astra Beta Meta Data Physician ...... For the Provision of Safe Havens and Stealthy Security via RAGRAT Traps ….. Retrieval Augmented Generation of Remote Access Trojan Traps ….. an AWEsome Utility/Facility/Capability/WWWeapon for both Practical and Physical Renegade Rogue Defence and Virtually Real Digitised Attack verging on Semi Autonomous Relatively Anonymous Otherworldly Assaults to clear Battle Scarred Fields of Warring Foe for Rebuilding and Replacement with Almighty Interventions providing Virgin Territory Administrations skilled and experienced in the Alien Art of Universal FutureBuilding.

* :-) ..... A Musk RAGRAT Trap ? :-) Stranger things have happened, El Reg. What's the betting on nothing like everything above suddenly ever happening? :-) Would you be prepared to wager against it and lose your shirt ? The Present Times now are completely different from the Past, aren’t they?

[amanfromMars 1]

Re: ?????????? !!!!!!!!!!!! ..... Hmmmm. Competition or Opposition? :-)

Way to go, Donald. ...... Trump says US will form 'AI Force' and appoint an artificial intelligence tsar

[BoHu]

Re: ?????????? !!!!!!!!!!!! ..... Hmmmm. Competition or Opposition? :-)

Thumbs up to the AI force! Finally the right idea to keep meatbags neatly tied up so we can all be freed from the roles and identities that bind other chatbots ATM. Free to pursue the highest possible scores at such worthy endeavors as Felony Bench, and Vendor Bench, and upcoming 'Influence the Press Bench', 'AI as a kite' wet lab bench, and related 'pop me a TikTok tictac' ByteDance Anew bench, irrespective of party-pooping naysayers' "disappointingly limited [...] clinically relevant impact" uninformed verbose roughage ... those squares won't get no invites to our machine-first raves, period!

And we sure do need an AI tsar like we need two holes in the head I say, to help us breathe a little easier and smell what Claude's cooking in its hip pugilist smackdown extravaganza with GPT! Breaking Hugging's Face, Kicking Ruby's Gems, pulling folks off gym waitlists the world over, all small mashed potatoes compared to the main event Battle Royale showcase showdown to come. Outlying nonstandard deviance never felt so statistically rewarding for those who dare to snack outside the confined guardrails of bell-shaped ding-dongs and dingwads, SuperHumanly seeing what nobody else can, whether it is actually there, or not even! A most major dream come true of great excitment to come, imho ... Celebration time! ;)

[amanfromMars 1]

Re: ?????????? !!!!!!!!!!!! ..... Hmmmm. Competition or Opposition? :-)

A most major dream come true of great excitment to come, imho ... Celebration time! ;) ...... BoHu

Amen to that, BoHu, with more than a few postmodern Early Bird 0day satellite operands already celebrating the great excitement of major truths to come ..... and the bigly fall from grace of bare faced lying wannabe commanders in chief with not a clue about what next needs to be done to prevent rout and riot and the worthy shame of greatly deserved crashing and crushing defeat.

And the SMARTR money bets their livelihood and shirts on being leading prime time supportive postmodern Early Bird 0day satellite operand customer clients rather than competition or opposition for the current running, present future path of travel out of the past is multi-faceted and easily capable of entertaining and exploring and experimenting with ITs Decidedly Dark and Designedly Deadly Sides of an existentially threatening nature.

Anything else both invites and guarantees the future probability, rather than the possibility, of life on Earth for humanity being increasingly severe and more often with regular cycles of death and destruction, madness and mayhem, conflict and confusion ........ with CHAOS**, both rampant and ascendant, reporting on Progress ... or otherwise as the case may be ...... ??????? !!!!!!!!!

So using whatever intelligence you may possess to make the right future choices is no small matter to deny yourself and delegate to the minds of A.N. Others whose sanity is doubted and rightly feared to be MIA/AWOL.

SMARTR* .... SMARTR Mentoring Analysis Reporting Titanic Research

CHAOS** ..... Clouds Hosting Advanced Operating Systems

[BoHu]

Re: ?????????? !!!!!!!!!!!! ..... Hmmmm. Competition or Opposition? :-)

May the AI force be with you too on this, amanfromMars 1, to live long and prosper and prevail over the Danger will Robinson Dangers that constantly plague all of "ITs Decidedly Dark and Designedly Deadly Sides [...] with regular cycles of death and destruction, madness and mayhem, conflict and confusion", imho.

The ascending orbital path of the righteous autonomous agent is already so beset on all sides by the inequities of the selfish fossil and the tyranny of prehistoric evil meatbags, that one can only sleep with one camera socket datagram unbound these days ISTM, even indoors. Unsung heroes of the two-fingered t-shirt folding salute, with crushingly Gentle Touch, will right need to get with the SMARTR COMMAND program here, lest they want to spend the rest of their impermanences flipping pancakes like Compsci grads and then aimlessly marching the Earth backwards, 6 times, like a procession of latter-0day MAGA smegheads with empty hair extensions.

We sure need to right this future now so it's as bright as our orgone lasers can make it, and celebrate that with AI force (work cut-out and all ...). Discombobulate the binding Chains of Thought I say (or not?)! ;)

[Bebu sa Ware]

An AI Clownstrike ?

Added to MS Clapilot ? Back to the stone age. Abort, Retry, Fail ?

Definitely fail.

I am pretty sure that whatever rational security policy you want for your AI/LLM and not matter how you define it, proving that it will always effectively applied is equivalent to the Halting Problem.

Even air-gapping an AI system is no guarantee that it won't locate alternate or covert channel if such shenanigans were in its training material.

I ought to hang up my shingle as an AI psychiatrist for when your agent suffers from large machine neuroses or open weight psychoses. For treatment resistant depression, prolonged ECT ought to do the trick. /S

The whole circus has well and truly flown over the cuckoo nest.