CVE-2025-48887 — MEDIUM (CVSS 6.5) AI Security Vulnerability

CISO Take

vLLM deployments with tool calling enabled are vulnerable to service disruption: any authenticated API user can send a crafted payload to trigger catastrophic regex backtracking in the tool call parser, taking down the inference service. Upgrade to vLLM 0.9.0 immediately; if patching is delayed, disable tool/function calling endpoints or add strict input length limits at the API gateway. Impact is limited to availability — no data exfiltration or code execution risk.

Risk Assessment

Operational risk is higher than CVSS 6.5 suggests for organizations running vLLM as a production inference endpoint. Attack complexity is trivial (crafted string input), requires only authenticated access (PR:L), and is network-reachable. The affected code path (pythonic_tool_parser.py) is active whenever function/tool calling is used — a standard pattern in agentic pipelines. EPSS of 0.00122 reflects low exploit-in-the-wild activity today, but the technique is well-understood and requires no ML expertise. Primary risk is availability of LLM inference infrastructure, not confidentiality.

Affected Systems

Package	Ecosystem	Vulnerable Range	Patched
vllm	pip	—	No patch
78.9K 126 dependents Pushed 6d ago 56% patched ~32d to patch Full package profile →
vllm	pip	>= 0.6.4, < 0.9.0	`0.9.0`
78.9K 126 dependents Pushed 6d ago 56% patched ~32d to patch Full package profile →

Severity & Risk

CVSS 3.1

6.5 / 10

EPSS

0.3%

chance of exploitation in 30 days

Higher than 57% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Trivial

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

Composite signal derived from CISA KEV, CISA SSVC, EPSS, trickest/cve, and Nuclei templates.

Attack Surface

AV Network

AC Low

PR Low

UI None

S Unchanged

C None

I None

A High

Recommended Action

5 steps

Patch

Upgrade vLLM to >= 0.9.0 (commit 4fc1bf813ad80172c1db31264beaef7d93fe0601 contains the fix).
Workaround

If immediate upgrade is blocked, disable the pythonic tool parser by switching to an alternative tool_call_parser (e.g., --tool-call-parser hermes) or disabling tool calling in your deployment config.
API gateway controls

Enforce maximum request payload size and per-user rate limits at the reverse proxy (nginx/Envoy) to reduce blast radius.
Detection

Alert on abnormal response latency spikes in your vLLM service metrics — ReDoS manifests as CPU saturation with near-zero throughput on the serving process.
Verify exposure

Check if your deployment enables --enable-auto-tool-choice or uses --tool-call-parser pythonic; those are the critical indicators.

CISA SSVC Assessment

Decision Track*

Exploitation poc

Automatable No

Technical Impact partial

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

Classification

DoS Inference API Framework AML.T0029 - Denial of AI Service AML.T0034 - Cost Harvesting AML.T0049 - Exploit Public-Facing Application

Compliance Impact

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2.6 - AI system availability and resilience

NIST AI RMF

MANAGE-2.2 - Risk treatment — availability and reliability

OWASP LLM Top 10

LLM04 - Model Denial of Service

Frequently Asked Questions

What is CVE-2025-48887?

vLLM deployments with tool calling enabled are vulnerable to service disruption: any authenticated API user can send a crafted payload to trigger catastrophic regex backtracking in the tool call parser, taking down the inference service. Upgrade to vLLM 0.9.0 immediately; if patching is delayed, disable tool/function calling endpoints or add strict input length limits at the API gateway. Impact is limited to availability — no data exfiltration or code execution risk.

Is CVE-2025-48887 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2025-48887, increasing the risk of exploitation.

How to fix CVE-2025-48887?

1. **Patch**: Upgrade vLLM to >= 0.9.0 (commit 4fc1bf813ad80172c1db31264beaef7d93fe0601 contains the fix). 2. **Workaround**: If immediate upgrade is blocked, disable the pythonic tool parser by switching to an alternative tool_call_parser (e.g., `--tool-call-parser hermes`) or disabling tool calling in your deployment config. 3. **API gateway controls**: Enforce maximum request payload size and per-user rate limits at the reverse proxy (nginx/Envoy) to reduce blast radius. 4. **Detection**: Alert on abnormal response latency spikes in your vLLM service metrics — ReDoS manifests as CPU saturation with near-zero throughput on the serving process. 5. **Verify exposure**: Check if your deployment enables `--enable-auto-tool-choice` or uses `--tool-call-parser pythonic`; those are the critical indicators.

What systems are affected by CVE-2025-48887?

This vulnerability affects the following AI/ML architecture patterns: LLM inference serving, Agent frameworks, Function calling / tool use endpoints, OpenAI-compatible API deployments, Multi-tenant AI platforms.

What is the CVSS score for CVE-2025-48887?

CVE-2025-48887 has a CVSS v3.1 base score of 6.5 (MEDIUM). The EPSS exploitation probability is 0.34%.

Technical Details

NVD Description

vLLM, an inference and serving engine for large language models (LLMs), has a Regular Expression Denial of Service (ReDoS) vulnerability in the file `vllm/entrypoints/openai/tool_parsers/pythonic_tool_parser.py` of versions 0.6.4 up to but excluding 0.9.0. The root cause is the use of a highly complex and nested regular expression for tool call detection, which can be exploited by an attacker to cause severe performance degradation or make the service unavailable. The pattern contains multiple nested quantifiers, optional groups, and inner repetitions which make it vulnerable to catastrophic backtracking. Version 0.9.0 contains a patch for the issue.

Exploitation Scenario

An adversary with legitimate API access to a vLLM-backed service (e.g., an internal AI platform user or a malicious external user of a public API) crafts a tool call payload with a pathological structure designed to trigger catastrophic backtracking in the nested quantifiers of the pythonic tool parser regex. The request is sent repeatedly via the OpenAI-compatible `/v1/chat/completions` endpoint. Each malicious request causes the regex engine to spin at near-100% CPU for an extended period. Within seconds of sustained requests, the vLLM serving process becomes unresponsive, denying service to all legitimate users. In agentic pipelines where the vLLM instance drives autonomous agents, this collapses the entire agent fleet — a single low-privilege user effectively disables production AI operations.