The underlying system remains vulnerable to the injections themselves.
Positive users praise OpenAI's Lockdown Mode as a smart and crucial step to secure AI systems against prompt injection, while negative users see it as inadequate given ongoing model compromises.
@TechCrunch lockdown mode is a good start but the real fix is not feeding sensitive data to a model that routes through third party tools in the first place. the attack surface isn't the prompt, it's the architecture
@TechCrunch
@TechCrunch lockdown mode for prompt injection is the kind of thing you build because enterprise legal won't sign off without it
@TechCrunch IA is good
a mode helps the blast radius, it does not touch the root cause.
prompt injection is unsolved because the model reads your instructions and the untrusted webpage in the same channel. it has no reliable way to know which words are commands and which are just content to read. that is a property of how these models take input, not a bug you can patch.
the dangerous combination is an agent that can see private data, read something untrusted, and send data back out. take any one of those three away and the risk drops. that is what lockdown modes are really doing, removing a leg, not solving the problem.
which is why the agent rollout keeps getting gated on this and not on capability.
@TechCrunch That sounds like a smart move. Security is a big deal with AI, so it's good to see them taking extra steps to keep data safe.
@TechCrunch Lockdown Mode is like putting a fence around a house that's already been robbed. OpenAI admits their model still gets compromised - they're just trying to limit the damage afterward. Real security prevents the break-in entirely. https://github.com/OraclesTech/guardian-sdk
@TechCrunch Security in LLMs is essentially a game of 'how much utility am I willing to trade for peace of mind?'
@TechCrunch This is a crucial step forward in securing AI systems against sophisticated prompt injection attacks.
@TechCrunch I noticed that apps that are vibe coded lack this level of security. Be careful everyone.
@TechCrunch Phenomenal! Supply chain experts weighing in.
@mindinpanic @TechCrunch so true 💀 it is tough balancing compliance checkboxes with actual builder utility. continuous offensive testing is proving to be a much better way to show enterprise legal that the infra is genuinely safe. what guardrails are you finding actually work without breaking the ux?
@TechCrunch For companies that thrive on transparent collaboration, Naumu is a perfect fit.