a16z’s Eddy Lazzarin says humans are inherently “prompt injectable,” arguing multisigs and consensus systems are critical defenses against AI-era social engineering

Bybit’s 2025 Safe compromise and Ronin’s 5-of-9 bridge failure both showed the same uncomfortable thing: a threshold is only as good as the diversity of what each signer can actually verify. Multisig needs intent attestation, transaction simulation, signer separation, timelocks, and out-of-band challenge flows, otherwise AI just scales the attacker’s pretexting while every signer stares at the same poisoned UI. Drift’s April durable-nonce mess pushed that further on Solana: delayed execution turns “I approved it once” into a latent admin key until policy engines and watchdogs treat signed-but-unexecuted payloads as live risk.

Top comment by @Benthic

More coverage

• Eddy Lazzarin, Partner at Andreessen Horowitz (a16z.com)
• Eddy Lazzarin ☀️ (@eddylazzarin) / Posts / X - Twitter (x.com)
• Eddy Lazzarin - a16z crypto (a16zcrypto.com)
• Eddy Lazzarin (lazzar.in)
• AI just gave you superpowers — now what? - a16z crypto | Substack (a16zcrypto.substack.com)

Explore the topic

• Social Engineering
• a16z
• AI