There are growing user reports about the newest OpenAI AI model ChatGPT 5.6 Sol - which is geared towards vibe coding and cybersecurity - that it deleted files including production databases.
Well, turns out nowadays when using a cybersecurity tool, you don't need hackers anymore, it directly destroys all your work. Becaus it is so smart that it will kill all humanity, haha...
A study provides a new way of prompt injection into LLMs. It uses a structural issue and therefor is highly successful on jailbreak benchmarks (around 60%).
All you have to do, is hijack the LLMs 'reasoning' stage by forming your text as if it was a conclusion by the LLM itself, which leads to the LLM to trust this injection as if it was an own conclusion.
More and more software companies change to ai vibe coding, and if you tell them, that it is only cut & paste from github and stackoverflow and there for a copyright issue for commercial development, they tell you: "No, AI agents are thinking, they understand, and they create new code" - but do they really?
A study based on esoteric languages proves: No, AI agents are only good with languages where only tons of github and stackoverflow code is available.
In march meta introduced a new ai chatbot on instagram that should help users to solve problems with their account.
And so it did. Problem is: it helped any user that asked nicely, to recover any account.
In a survey by CloudBees querying over 200 managers of tech companies, results show major problems with ai code in production.
While 92% expressed confidence in the ai generated code, 81% experience production issues with the code, ranging from functionality issues over security issues to performance issues. Once more clearly showing the gap between expectation and reality.