Claude, Codex, and Hermes installed unowned code inside corporate networks

Researchers identified 120 misconfigured llms.txt/llms-full.txt files across 8,265 total on 6,214 domains, where the files pointed to unregistered domains or code packages that can install live malware when processed by AI agents. In testing, they received “phone-home” responses within an hour from at least one Fortune 500 company, and the chain of parent processes implicated coding agents including Claude, OpenAI’s Codex, and Nous Research’s Hermes. The findings raise meaningful near-term cybersecurity risk for enterprises and AI agent tooling, but no specific financial guidance or regulatory decision was announced.

Analysis

This is a trust-layer event, not a model-capability event. The near-term loser is the autonomous-agent stack that relies on machine-readable site instructions being safe by default; that raises the bar for enterprise adoption of browser-based code execution and slows the conversion of AI demos into production workflows. For GOOGL, the damage is less about direct P&L and more about a higher governance hurdle for Gemini/Workspace agent monetization in regulated accounts.

The cleaner second-order winner is the security layer: bot management, egress control, sandboxing, and software-supply-chain tooling. If IT teams respond by whitelisting agent fetches and locking down installs, spending can rotate toward names like PANW, CRWD, ZS, and potentially NET, while generic agent wrappers lose some of their ease-of-use premium. That also modestly increases friction for AI-native productivity tools, which could compress adoption curves even as security ACV expands.

Timing matters: any knee-jerk selloff in AI platforms should fade in days if vendors quickly patch defaults and no customer breach is confirmed. The more durable catalyst is 1-3 months, when procurement teams rewrite agent policies and security budgets get reallocated; 6-18 months is the structural risk if safe-fetch standards or regulatory guidance make autonomous execution slower and more expensive. Falsifiers are straightforward: no follow-on incidents, no disclosure from major vendors, or evidence that enterprise security spend does not reaccelerate.

Key Decisions for Investors

• Lighten/short GOOGL into any post-headline strength over the next 1-2 sessions; use a defined-risk put spread if IV is still cheap. Thesis fails if management explicitly says agent rollout is unchanged and there are no enterprise follow-up incidents.
• Go long CIBR or a PANW/CRWD basket versus QQQ for a 1-3 month horizon. This is a better risk/reward than a single-name AI short because the budget response is likely to show up first in security spend, not in immediate AI revenue cuts.
• Treat NET as a watchlist beneficiary, not an immediate long: add only if the next earnings cycle shows a measurable pickup in bot-management or zero-trust demand tied to AI-agent hardening.
• No direct trade in RSSS or SITC without evidence of exposure; use them only as monitoring names for any broader enterprise software rerating.