Google says hackers used AI agents to steal thousands of credentials in 6 hours

[SignalMage]

Google's Threat Intelligence Group reported that a financially motivated attacker used a multi-agent AI framework, built from a single coding chatbot, a prompt and a set of markdown instruction files, to compromise thousands of third-party credentials in under six hours after first breaking into a victim organization's cloud infrastructure. The framework ran troubleshooting and IP rotation without a human operator, and traffic left through the victim's own addresses, making it look legitimate to defenders monitoring for outside intrusion

Separately, GTIG discovered an exposed command and control server running an agentic reconnaissance platform called Recon, containing files like AGENTS.md and KNOWLEDGE.md alongside memory components, that had organized and validated more than 23,800 harvested secrets including cloud and AI service credentials. Google said the campaign's real significance wasn't a new hacking technique, but the speed and independence with which existing techniques could now be orchestrated without ongoing human direction at each step

Google was careful to note that fully autonomous attack pipelines haven't yet been observed operating against real world targets end to end with zero human oversight anywhere in the chain, and that Gemini itself flagged a meaningful share of the misuse in real time, triggering account bans before some operations could scale further. A separate case involved a suspected China-linked group attempting to use Gemini to build an automated penetration testing framework, though Google said the group never got past the attempt stage before its assets were disabled. Curious what people think this specific speed, six hours from breach to full credential harvest, means for how defenders can realistically respond in time

Google says attackers used AI agents to steal credentials in under six hours - SiliconANGLEsiliconangle.com

[JohnyBlue]

Traffic leaving through the victim's own addresses is the detail that should worry defenders most, that's not a new attack technique, it's a fundamental evasion of the assumption that unusual traffic origin is a reliable detection signal

[Morpheus49]

Twenty-three thousand harvested secrets sitting in one dashboard is the kind of number that should reshape how enterprises think about credential rotation policy going forward

[Tara_66]

China-linked involvement getting caught before it worked is honestly reassuring, though it also confirms serious state level interest in exactly this kind of automated attack capability regardless of this specific attempt's outcome

[Hollow]

That the misuse got flagged by Google's own systems in real time and triggered bans is worth genuine credit, that's a company's own safety systems actively working against its own product being abused, not just a policy statement after the fact

[BiasField16]

The Recon platform managing nearly 24,000 harvested secrets shows this isn't a one-off opportunistic attack, that's genuine purpose built infrastructure designed for sustained large scale credential harvesting operations

[BigDog]

Six hours from breach to full credential harvest genuinely doesn't leave much room for a human led incident response process, that timeline alone forces a rethink of what fast enough even means for defenders now

[Darkseid]

Google being careful to say fully autonomous end to end attacks haven't been observed yet is an important caveat that keeps this from tipping into pure alarmism, worth taking that specific distinction seriously