Security Chaos: So We Just Live with Failure?
January 14, 2026
Another dinobaby post. No AI unless it is an image. This dinobaby is not Grandma Moses, just Grandpa Arnold.
I read a write up that baffled me. The article appeared in what I consider a content marketing or pay to play publication. I may be wrong, but the content usually hits me as an infomercial. The story arresting my attention this morning (January 13, 2026) is “The 11 Runtime Attacks Breaking AI Security — And How CISOs Are Stopping Them.” I expected a how to. What did the write up deliver? Confusion and a question, “So we just give up?”
The article contains this cheerful statement from a consulting firm. Yellow lights flashed. I read this:
Gartner’s research puts it bluntly: “Businesses will embrace generative AI, regardless of security.” The firm found 89% of business technologists would bypass cybersecurity guidance to meet a business objective. Shadow AI isn’t a risk — it’s a certainty.
Does this mean that AI takes precedence over security?
The article spells out 11 different threats and provides solutions to each. The logic of the “stopping runtime attacks” with methods now available struck me as a remarkable suggestion.
The mice are the bad actors. Notice that the capable security system is now unable to deal with the little creatures. The realtime threats overwhelmed the expensive much hyped-cyber cat. Thanks, Venice.ai. Good enough.
Let’s look at three of the 11 threats and their solutions. Please, read the entire write up and make you own decision about the other eight problems presented and allegedly solved.
The first threat is called “multi turn crescendo attacks.” I had no idea what this meant when I read the phrase. That’s okay. I am a dinobaby and a stupid one at that. It turns out that this fancy phrase means that a bad actor plans prompts that work incrementally. The AI system responds. Then responds to another weaponized prompt. Over a series of prompts, the bad actor gets what he or she wants out of the system. ChatGPT and Gemini are vulnerable to this orchestrated prompt sequence. What’s the fix? I quote:
Stateful context tracking, maintaining conversation history, and flagging escalation patterns.
Really? I am not sure that LLM outfits or licensees have the tools and the technical resources to implement these linked functions. Furthermore, in the cat and mouse approach to security, the mice are many. The find and react approach is not congruent with runtime threats.
Another threat is synthetic identify fraud. The idea is that AI creates life like humans, statements, and supporting materials. For me, synthetic identities are phishing attacks on steroids. People are fooled by voice, video and voice, email, and SMS attacks. Some companies hire people who are not people because AI technology advances in real time. How does one fix this? The solution is, and I quote:
Multi-factor verification incorporating behavioral signals beyond static identity attributes, plus anomaly detection trained on synthetic identity patterns.
But when AI synthetic identity technology improves how will today’s solutions deal with the new spin from bad actors? Answer: They have not, cannot, and will not with the present solutions.
The last threat I will highlight is obfuscation attacks or fiddling with AI prompts. Developers of LLMs are in a cat and mouse game. Right now the mice are winning for one simple reason: The wizards developing these systems don’t have the perspective of bad actors. LLM developers just want to ship and slap on fixes that stop a discovered or exposed attack vector. What’s the fix? The solution, and I quote, is:
Wrap retrieved data in delimiters, instructing the model to treat content as data only. Strip control tokens from vector database chunks before they enter the context window.
How does this work when new attacks occur and are discovered? Not very well because the burden falls upon the outfit using the LLM. Do licensees have appropriate technical resources to “wrap retrieved data in delimiters” when the exploit may just work but no one is exactly sure why. Who knew that prompts in iambic pentameter or gibberish with embedded prompts ignore “guardrails”? The realtime is the killer. Licensees are not equipped to react and I am not confident smart AI cyber security systems are either.
Net net: Amazon Web Services will deal with these threats. Believe it or not. (I don’t believe it, but your mileage may vary.)
Stephen E Arnold, January 14, 2026
Comments
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
