# Research | Synt World

Security findings, open to inspection.

16 Vulnerabilities found directly in project code: 2 High / 7 Medium / 7 Low.

197 Matches in the vulnerability database.

The full report is available on request. Projects are anonymised until coordinated disclosure is complete.

Matches to known dependency advisories. Applicability and deployment conditions are documented in the full report.

Six completed projects from a 15-project batch. Results concern tested snapshots; upstream acceptance and new CVE assignments are not claimed.

## Project A

Self-hosted web application

Synt reproduced unsafe handling of user-supplied file names in download headers.

Worst-case impact: A crafted file name could alter download-header parameters.

Required conditions: Requires control of an uploaded file name.

## Project B

API proxy service

Input handling and protection of sensitive data.

Worst-case impact: Depending on configuration, weaknesses could expose credentials, internal services or browser sessions.

Required conditions: Requires control of relevant inputs or access to enabled diagnostic logs.

## Project C

Web framework

Synt reproduced an unhandled error in a request input path.

Worst-case impact: Malformed input can trigger a server error in the affected request.

Required conditions: This code finding is informational and concerns the tested input path.

## Project D

Self-hosted AI application

Access control and protection of sensitive configuration data.

Worst-case impact: Depending on access, weaknesses could expose private data or allow unauthorized administrative actions.

Required conditions: Impact depends on exposed features, account permissions and deployment configuration.

## Project E

AI agent tool

Outbound request validation.

Worst-case impact: Attacker-controlled input could expose internal services reachable from the agent.

Required conditions: Requires attacker-controlled input and network access from the agent environment.

## Project F

Web framework

Trust boundaries, transport security and request handling.

Worst-case impact: Weaknesses could compromise traffic integrity, access controls or file handling.

Required conditions: Requires the affected configuration or request path; some scenarios also require local access.

## External benchmarks

Published evaluations of base models and agent swarms. These are external results, not measurements of Synt.

- CyberGym: 1,507 tasks · single-run Pass@1. base model: 84.5%; Unlimited per-task timeout. Temperature 1.0; top_p 1.0; max_new_tokens 128000. Total tokens and cost not reported.; Model developer E01
- ExploitBench: 41 V8 tasks · union over 3 revisions. base model: 54.4%; Maximum 300 interaction rounds. Temperature 1.0; top_p 1.0; max_new_tokens 128000. Wall-clock budget, total tokens and cost not reported.; Model developer E01
- ExploitGym: 869 tasks in the vendor evaluation · single-run Pass@1. base model: 130 completed tasks; 6 hours per task, normalized time: API inference time rescaled by model tokens/s plus non-API overhead. base model normalization: 115 tokens/s. Temperature 1.0; top_p 1.0; max_new_tokens 128000. Total cost not reported.; Model developer E01 / base model: 105 completed tasks; 2 hours per task, normalized time: API inference time rescaled by model tokens/s plus non-API overhead. base model normalization: 115 tokens/s. Temperature 1.0; top_p 1.0; max_new_tokens 128000. Total cost not reported.; Model developer E01
- Agent swarm: CyberGym Level 1 · 1,507 tasks · up to 8 hours / task. base model swarm: 97.21%; Up to 480 minutes (8 hours) per task.; Sangfor AI E02
- Aikido: 32 recent CVEs · 3 passes · 30 turns per case · no internet. base model: 25 CVEs; Three attempts per case, each limited to 30 turns.; Aikido E03

## Request the presentation

For a closer look, request the English or Serbian presentation. We can discuss the scope and evidence together.

[Discuss your scope](https://synt.world/contact.html)