OWASP_LLM_TOP10LLM02:2025voice-validated

OWASP_LLM_TOP10 LLM02: LLM02:2025

OWASP_LLM_TOP10

AL
Adam Lundqvist
Founder at SQUR · last verified 2026-06-20

Regulation text

LLMs can expose sensitive information through their outputs, including personally identifiable information (PII), proprietary algorithms, confidential business data, training data leakage, and credentials. Risks include lack of input/output sanitisation, inadequate data anonymisation, system prompt leakage, and indirect inference attacks that probe the model for sensitive context.

ATT&CK techniques this article tests · 15

TechniqueWhy it mapsConfidence
T1598.0031. Attackers craft malicious prompts, exploiting LLM vulnerabilities to elicit sensitive data, including PII and proprietary information.
90%
T1592.0011. LLMs, when improperly secured, can inadvertently disclose details about the underlying host hardware, aiding reconnaissance efforts.
80%
T1595.0021. LLMs trained on or given access to confidential business documents may reveal strategic plans or competitive intelligence through crafted queries.
85%
1. LLMs with access to local file systems can be prompted to read and output sensitive data stored on the system.
90%
T1074.0011. LLM outputs containing sensitive information effectively stage data for collection, simplifying exfiltration for attackers.
85%
1. Automated data processing by LLMs, if not properly controlled, can lead to the unintended collection and exposure of sensitive information.
80%
1. LLMs integrated with cloud storage can be manipulated to retrieve and disclose sensitive data from cloud objects.
85%
T1560.0011. Attackers can prompt LLMs to summarize or archive sensitive data, consolidating it for easier exfiltration.
75%
1. Sensitive data extracted from an LLM via prompt injection can be considered exfiltrated over a C2 channel, as the attacker controls the output mechanism.
90%
T1567.0021. An LLM can be coerced into outputting sensitive data to an attacker-controlled web service or cloud storage, bypassing traditional network controls.
85%
T1552.0011. LLMs may inadvertently reveal credentials if they are present in training data, accessible files, or system prompts.
90%
1. LLMs can disclose internal file paths, directory structures, or sensitive file names, aiding attacker discovery.
80%
T1087.0011. LLMs, through inference or direct access, can reveal information about user accounts or system accounts.
75%
1. LLMs can be prompted to disclose details about internal network systems, connected services, or network topology.
70%
T1027.0061. LLMs can be used to embed sensitive information within seemingly innocuous outputs or extract information hidden in inputs, acting as a steganographic tool.
65%

Defending mitigations · 6

MitigationWhat it doesConfidence
M10311. Network segmentation isolates LLM systems from sensitive internal networks, preventing unauthorized data access and disclosure.
90%
M10351. Limiting LLM access to only necessary resources prevents it from accessing and disclosing sensitive data it does not require for its function.
95%
M10381. Implementing robust user account management ensures LLM services operate with the principle of least privilege, reducing potential data exposure.
85%
M10401. Data Loss Prevention (DLP) solutions monitor LLM outputs, detecting and blocking the disclosure of sensitive information like PII or proprietary data.
90%
M10471. Comprehensive auditing of LLM inputs and outputs provides visibility into potential sensitive information disclosure attempts and actual leaks.
85%
M10501. Regular vulnerability scanning identifies weaknesses in LLM applications and their integrations that could lead to sensitive data exposure.
80%

Underlying weaknesses · 7

CWEWhy it persistsConfidence
CWE-201. Improper input validation allows malicious prompts to bypass security measures, leading to sensitive information disclosure.
95%
CWE-2001. The core vulnerability: LLMs expose sensitive information (PII, proprietary data) to unauthorized users through their outputs.
98%
CWE-3591. LLMs can violate privacy by disclosing private information, including PII, due to inadequate data anonymization or training data leakage.
90%
CWE-5321. Sensitive information from prompts or internal processing may be inadvertently included in LLM log files, creating a disclosure risk.
85%
CWE-5401. System prompt leakage occurs when sensitive configuration details or hardcoded secrets are exposed within the LLM's accessible components.
80%
CWE-6681. LLMs with excessive permissions or misconfigurations can access resources (e.g., internal databases) outside their intended scope, leading to data exposure.
85%
CWE-9181. An LLM vulnerable to SSRF can be tricked into making requests to internal services, potentially revealing sensitive internal network information.
75%

What SQUR Covers

Web application + API pentesting for OWASP Top 10, business logic flaws, authentication bypass, injection attacks, and other application-layer vulnerabilities. €1,995 per scan, 24-hour turnaround, EU-only data.

What SQUR Does Not Cover

Internal network pentesting, endpoint security testing, physical security assessments, social engineering, or ICT third-party concentration risk reviews. Engage a complementary provider for those scope items.

Provenance

Mapped Q2.2026 using gemini-2.5-flash · €0.0183 compute · voice-rubric self-validated