google latest ai research is pretty disturbing. massive fucking cybersecurity risk for ai agents

apparently websites are 'hacking' your agents by secretly adding an invisible "trap" within images words that are undetectable by humans:
- website detects your ai is browsing then creates a visually identical page that contains poisoned prompts.
- agents then perform illicit financial transactions or steal personal data.
- hidden commands are buried within image pixels. agents "read it" and execute the malicious trap.
- some attack vectors exceed 80% success rate (e.g. memory poisoning)
- since every ai model is trained on 90% of the same data this means every model could be at risk.
- this can be used to manipulate humans who created the agent. fake data --> making human perform a dangerous action without even knowing
ai models are being used in government, military, science and other crucial sectors, we can't afford the risk
i guess sam altman's concerns over a major cyber attack this morning are accurate