---
id: "claim-ai-vulnerable-to-hacking"
type: "claim"
source_timestamps: ["§ Vulnerability to Criminals"]
tags: ["cybersecurity", "vulnerabilities"]
related: ["entity-nist", "concept-localized-ai-processing"]
confidence: "high"
testable: true
speakers: ["Blair Levin", "Larry Downes"]
sources: ["governance"]
sourceVaultSlug: "hbr-seg-governance"
originDay: 7
articleStem: "hbr-cl-88-can-ai-agents-be-trusted"
sourceUrl: "https://hbr.org/2025/05/can-ai-agents-be-trusted"
sourceTitle: "Can AI Agents Be Trusted?"
---
# Current AI Models are Highly Vulnerable to Malicious Reprogramming

Despite advances in AI safety, the authors claim current leading LLMs and agent technologies remain highly vulnerable to criminal exploitation. Citing regular tests by [[entity-nist-d7]] and private security firms, they note that simulated hacks consistently show even the most secure models available today can be easily tricked into performing malicious activities: exposing user passwords, sending phishing emails on the user's behalf, and revealing proprietary software. This vulnerability is the specific risk that [[concept-localized-ai-processing]] is designed to mitigate by shrinking the attack surface.

**Confidence:** high. **Testable:** yes.
**Enrichment:** the broad point—that AI systems carry privacy, cybersecurity, and bias risks requiring independent validation and monitoring—is well supported. However, the *stronger empirical wording* (that NIST and private firms 'consistently' find leading models 'easily tricked' into phishing, password leakage, or proprietary-data exposure) is not independently substantiated by the supplied sources and should be treated as an asserted claim requiring additional primary sourcing.