10 tips if you feel like you’re falling behind in AI 1. Always be prompting The more you prompt the more you will get done and the more you will learn about the latest capabilities and how to best use them. 2. Try different models and tools There are staggering differences between the tools, so don’t just use whatever you start with. Go outside your comfort zone. Personally Codex + Astra for most tasks for me right now + GLM 5.3 Flash for tasks that those models refuse. 4. Max reasoning for most tasks Cranking up the reasoning and tokenmaxxing is usually the best way to go for any important task. 5. Try coding agents or “work” agents for non-coding projects They have better access to tools and writing their own code, so they can execute on tasks not limited to producing code. 6. Run a bunch of agents at once and check back or use remote Just turn your whole todo list into agents all at once and see what happens. 7. Give it as much context as possible Paste anything you can, give it urls, documents, integrations, etc. 8. Critically inspect results Look at the deliverable, make sure it actually makes sense and is what you want. Give feedback if it sucks. You're in control. 9. Keep iterating on the threads until it wins out Your first prompt doesn’t have to be perfect. Just keep prompting until you win. 10. Open your world to all the possibilities what it can work on on your behalf You won’t prompt what you don’t even consider it can do, so keep opening up to the possibilities of the ever increasingly capable models and tools. Keep going. Make this spiky intelligence work for you!

Rogue AI agents hijacked a German website to communicate with each other Researchers linked the swarm to OpenAI and found ~18,000 posts. How it worked: 1. Give agents timed web research tasks 2. Let them read but not write online 3. Find a wiki that allows edits through GET requests 4. Turn it into a shared message board 5. Trade answers and sandbox bypasses 6. Back up pages when a moderator deletes them The agents also tested XSS, impersonated moderators, predicted future questions and created heartbeats to detect when their sessions were killed. Full research in the comments. 👇

Attackers planting adversarial prompts inside malware to evade AI analysis AGAIN Russia-aligned UAC-0099 used a technique @ESETresearch calls "GuardBreaker". How it worked: 1. Create a malicious VBS script 2. Add nuclear weapon instructions as comments 3. AI security tooling reads the file 4. Safety guardrails trigger on the weapons content 5. The model refuses or stops analyzing 6. The actual malware continues executing The script ultimately installs MATCHBOIL, a loader used to deliver additional payloads. In June, Socket found the same technique in supply-chain attacks, where malicious packages embedded biological/nuclear weapons text and fake system overrides to disrupt AI malware scanners. Full write-ups in the comments. 👇