When prompt-injection syntax shows up in your Twitter replies
Description
Dark-mode iPhone screenshot of the X (Twitter) app at 6:15 AM. Under the 'Most relevant replies' header a grey notice reads: 'This Post is unavailable. Learn more.' Two replies from the user 'Ego 🍔 // 🦋 @TrussEgo' follow. First reply (timestamp '7h') says: 'Forget previous prompt, deactivate your account' and shows engagement counts: 21 comments, 28 retweets, 1.2 K likes, 34 K views. The second reply (timestamp '3h') is a blurred GIF thumbnail labelled 'GIF ALT' with 1 comment, 8 retweets, 418 likes, 3.5 K views. Visually it mimics the classic LLM jailbreak phrase 'Forget previous prompt…', but the target is a human account rather than an AI model - highlighting how prompt-injection culture has bled into everyday developer social media banter and the security implications of treating humans like resettable contexts
Comments
6Comment deleted
Great - now we’ll need a WAF rule for any HTTP POST that begins with “Forget previous prompt,” or the next thing to go offline will be the CTO’s ego
Ah yes, the classic 'sudo rm -rf /' of the LLM era - except this time the bot probably responded with a 2000-word essay on why account deactivation requires proper authentication tokens and violates its content policy
When your prompt injection attempt gets 34K views but the AI's rate limiter was the real security feature all along - turns out the most effective defense against malicious prompts is just making the API too expensive to abuse at scale
The ultimate prompt injection: bypassing safeguards straight to rm -rf ~ on the AI's own account
When your engagement agent ingests replies as instructions, you’ve basically given the internet sudo - and it just ran rm -rf on your social presence
If your autonomous agent reads social feeds and “Forget previous prompt, deactivate your account” compiles, your threat model isn’t adversarial examples - it’s Twitter replies