When LLM Defenses Backfire: Characterizing Safety, Performance, and Cost Trade-Offs

Jailbreak defenses are essential for protecting large language models LLMs, but they can also introduce secondary costs that weaken model utility. We present a systematic study of these defense trade-offs along three dimensions: performance impact, over-refusal on benign inputs, and inference cos...

CVE-2026-66005

CVE-2026-66005 affects Jan up to version 0.8.4. The local API server has a CORS misconfiguration that replaces user-configured trusted hosts with a wildcard, reflecting arbitrary origins with credentials. This enables network-adjacent attackers to bypass trusted host restrictions and reach the un...

CVE-2026-66005

Jan through 0.8.4, fixed in commit 3e1c1e7, contains a CORS misconfiguration vulnerability in its local API server that allows network-adjacent attackers to bypass trusted host restrictions by exploiting the server's replacement of user-configured trusted hosts with a wildcard that reflects...

PT-2026-64318

Name of the Vulnerable Software and Affected Versions Jan versions prior to 0.8.4 Description A CORS Cross-Origin Resource Sharing misconfiguration in the local API server allows network-adjacent attackers to bypass trusted host restrictions. This occurs because the server replaces user-configure...

caption-flow (>=0.5.0, <=0.5.1), lazyllm-verl (0.3.2.dev1) +3 more potentially affected by CVE-2026-54234 via vllm (>=0.19.0, <=0.23.0)

vllm PYPI version =0.19.0, =0.5.0, =2.5.3, =0.1.1.4, =2.1.2, =2.2.4 Source cves: CVE-2026-54234 Source advisory: OSV:PYSEC-2026-3542...

CVE-2026-65701: Improper Limitation of a Pathname to a Restricted Directory

SoftVC VITS Singing Voice Conversion through commit 730930d contains a path traversal vulnerability in the full-song inference server that allows unauthenticated remote attackers to read and exfiltrate arbitrary files by supplying attacker-controlled filesystem paths through the audiopath field o...

Docker Desktop for macOS Inference Server Permissive Allow List Sandbox Escape Vulnerability

This vulnerability allows local attackers to escape the model runner sandbox on affected installations of Docker Desktop for macOS. An attacker must first obtain the ability to execute low-privileged code within the sandbox in order to exploit this vulnerability. The specific flaw exists within t...

EUVD-2026-41925

vLLM denial of service via prompt embeds on M-RoPE models...

mlopspython-inference (>=67580760626172675.0.0, <=757069802391466481559610331145) potentially affected by CVE-2026-12484 via keras (2.12.0rc0)

keras PYPI version 2.12.0rc0 is affected by a known vulnerability. The following packages have a transitive dependency on keras and may be impacted: - mlopspython-inference =67580760626172675.0.0, =757069802391466481559610331145 Source cves: CVE-2026-12484 Source advisory: OSV:GHSA-V2W2-W228-C444...

caption-flow (>=0.5.0, <=0.5.1), lazyllm-verl (0.3.2.dev1) +3 more potentially affected by CVE-2026-54234 via vllm (>=0.19.0, <=0.23.0)

vllm PYPI version =0.19.0, =0.5.0, =2.5.3, =0.1.1.4, =2.1.2, =2.2.4 Source cves: CVE-2026-54234 Source advisory: OSV:GHSA-8WR5-JM2H-8R4F...

caption-flow (>=0.5.0, <=0.5.1), lazyllm-verl (0.3.2.dev1) +3 more potentially affected by CVE-2026-54234 via vllm (>=0.19.0, <=0.23.0)

vllm PYPI version =0.19.0, =0.5.0, =2.5.3, =0.1.1.4, =2.1.2, =2.2.4 Source cves: CVE-2026-54234 Source advisory: GHSA-8WR5-JM2H-8R4F...

Code-Poisoning Property Inference Attacks

The flourishing code hosting platforms and coding agents enable even beginners with private data to build tailored Machine Learning ML models using available code quickly. The training data for ML models, often regarded as private property e.g., clinical records, transaction information, is at...

CVE-2026-63086

text-generation-inference through 3.3.7 contains a server-side request forgery SSRF vulnerability in the OpenAI-compatible multimodal chat completions endpoint that allows unauthenticated network attackers to coerce the server into issuing arbitrary HTTP GET requests by supplying a crafted imageu...

CVE-2026-63086

The CVE-2026-63086 entry concerns text-generation-inference 3.3.7, with an SSRF flaw in the OpenAI-compatible multimodal chat completions endpoint. The fetch_image function in router/src/validation.rs does not validate addresses (private, loopback, link-local, or cloud metadata). The HTTP client ...

CVE-2026-63086

text-generation-inference through 3.3.7 contains a server-side request forgery SSRF vulnerability in the OpenAI-compatible multimodal chat completions endpoint that allows unauthenticated network attackers to coerce the server into issuing arbitrary HTTP GET requests by supplying a crafted imageu...

CVE-2026-63086 text-generation-inference 3.3.7 SSRF via fetch_image in multimodal chat completions

text-generation-inference through 3.3.7 contains a server-side request forgery SSRF vulnerability in the OpenAI-compatible multimodal chat completions endpoint that allows unauthenticated network attackers to coerce the server into issuing arbitrary HTTP GET requests by supplying a crafted imageu...

Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents

Security-agent evaluations commonly measure peak offensive capability under generous inference budgets, emphasizing vulnerability discovery, exploit development, penetration testing, and CTF completion. Such measurements are useful but incomplete: in operational security, every reasoning step, to...

CVE-2026-63086: Server-Side Request Forgery (SSRF)

text-generation-inference through 3.3.7 contains a server-side request forgery SSRF vulnerability in the OpenAI-compatible multimodal chat completions endpoint that allows unauthenticated network attackers to coerce the server into issuing arbitrary HTTP GET requests by supplying a crafted imageu...

CVE-2026-58658

GPUStack through 2.2.1, fixed in commit 4e20551, contains an unauthenticated information disclosure vulnerability that allows unauthenticated attackers to access sensitive inference logs and modify worker configuration by exploiting unprotected /serveLogs and /debug endpoints on the worker port...

CVE-2026-58658 GPUStack Unauthenticated Information Disclosure via Worker Endpoints

GPUStack through 2.2.1, fixed in commit 4e20551, contains an unauthenticated information disclosure vulnerability that allows unauthenticated attackers to access sensitive inference logs and modify worker configuration by exploiting unprotected /serveLogs and /debug endpoints on the worker port...