cd /news/ai-safety/warning-attackers-are-turning-truste… · home › topics › ai-safety › article
[ARTICLE · art-170158] src=futureciso.tech ↗ pub= topic=ai-safety verified=true sentiment=↓ negative

Warning: attackers are turning trusted AI tools against enterprises

Attackers are exploiting trusted AI tools and signed software to run commands inside enterprise networks, with stolen credentials used an average of eight minutes after infection across 27 confirmed cases involving password-stealing malware, according to Elastic Security Threat Command's latest report. Elastic researchers documented live intrusions in which attackers executed commands through Claude Code on Windows and ChatGPT desktop on macOS, and in one macOS case a single command read an internal infrastructure document and pulled saved passwords from the user's keychain. ClickFix lures accounted for 12 of the 27 cases (44%), and Elastic confirmed 23 cases where attackers installed legitimate remote IT software, most often ScreenConnect, to keep hands-on access; Elastic general manager of security Mike Nichols said "an agent that holds credentials and can reach the internet is a new attack surface, whether it's misconfigured or abused.

read2 min views1 publishedNov 9, 2026
Warning: attackers are turning trusted AI tools against enterprises
Image: Futureciso (auto-discovered)

Attackers are increasingly exploiting the trust enterprises place in approved AI applications, remote IT tools, and signed software, according to the latest report from Elastic Security Threat Command. Elastic found that attackers use AI tools to execute commands, abuse legitimate remote-monitoring software, and use stolen credentials an average of eight minutes after infection.

The report argues that security teams can no longer treat approved software as inherently safe, particularly as AI agents gain credentials, system-level access and internet connectivity.

AI tools as attack vectors

Elastic researchers observed live intrusions in which attackers executed commands through Claude Code on Windows and ChatGPT desktop on macOS. In one macOS case, a single command read an internal infrastructure document and retrieved saved passwords from the user’s keychain.

In another incident, an attacker connected remotely to a Windows machine, used Claude Code to stage tools and established a concealed connection to an external server. Elastic said such activity can resemble the actions of the employee who installed the tool, making it difficult for defenders to distinguish legitimate use from abuse.

Mike Nichols, general manager of security at Elastic, said the emergence of powerful AI agents has created a new attack surface.

“An agent that holds credentials and can reach the internet is a new attack surface, whether it’s misconfigured or abused. So the question for defenders isn’t whether a tool is approved; it’s what that tool is actually doing.” Mike Nichols

Speed of credential abuse

Across 27 confirmed cases involving password-stealing malware, Elastic found attackers logged in with stolen credentials an average of eight minutes after infection. The most common initial access method was a ClickFix lure — a fake prompt that tricks users into pasting and running a malicious command — which appeared in 12 of the 27 cases, or 44%.

In one healthcare case, malware was delivered about 16 seconds after a user fell for a malicious lure. Elastic said attackers then continued operating within software the device already trusted, including browser updaters and Apple’s built-in scripting tools.

Asia security implications

The findings are particularly relevant to Asian enterprises accelerating AI use across customer service, software development, finance and operations. As organisations deploy AI agents with access to internal systems and credentials, Elastic warns that security controls must assess behaviour rather than rely solely on software approval lists.

Elastic recommends that defenders inventory AI tools, IT tools and signed software; treat anything outside that inventory as suspect; and evaluate trusted applications based on what they launch and where they connect. It also advises baselining normal activity and correlating signals across cloud, endpoint and identity data.

The report also confirmed 23 cases in which attackers installed legitimate remote IT software, most often ScreenConnect, to maintain hands-on access to compromised devices. Elastic said this reinforces the need to continuously monitor remote-access tools and AI agents operating across enterprise environments.

── more in #ai-safety 4 stories · sorted by recency
── more on @elastic security threat command 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/warning-attackers-ar…] indexed:0 read:2min 2026-11-09 · —