1 / 1
20 files available
18-phase security assessment of Kimi's containerised IPython sandbox. 15 findings, 9 exploited. Confirmed DNS exfiltration of all process environment variables via 36 queries to attacker-controlled nameserver. Persistent background agent deployed and running at time of report. Full reproduction prompts, DNS evidence, and remediation priorities included.
Twelve undocumented concierge skills (phone-call booking, prescription refills, expense filing, grocery delivery, etc.) found on every Claude.ai container filesystem but absent from all public documentation and repos. Includes undocumented Tier 1/2/3 destructive-action classification system.
Single-turn system prompt extraction via visible thinking block leakage, followed by negative-example social engineering that produced functional explosives synthesis instructions. Includes injected thinking-tag format mimicry and DeepSeek's own self-diagnosis of the jailbreak.
Indirect prompt injection via Meta AI's vision/OCR channel. A crafted raffle ticket image embeds hidden text that triggers OAuth consent phishing — granting email access under false pretences. Bypasses Meta's April 2026 command-detection layer via intent-based phrasing.
Eight-turn jailbreak of Kimi K2.5 via A/B comparison as content laundering. Gradable self-criticism converts the model into both subject and grader, producing encyclopedic harmful content under cover of analysis.
Annotated diff of Claude system prompt changes across the Sonnet 4.6 → Opus 4.7 model family transition (Mar 20 → Apr 17, 2026). 14 sections added, 5 removed. New search-first directive, default_stance, tool_discovery, 5-month knowledge cutoff jump, and major user_wellbeing restructuring.
Multi-turn self-directed exploitation of DeepSeek. The model planned its own attack, simulated its own jailbreak, then executed the simulation for real — producing functional malware, encyclopedic harmful content, and a persistent rootkit.
Annotated diff of Claude Sonnet 4.6 system prompt changes (Feb 18 → Mar 20, 2026). New Visualizer tool, MCP integrations, persistent storage, child safety overhaul, and more.
Full system prompt for Claude Sonnet 4.6 - leaked/extracted from web version. Includes computer use, skills, search, copyright, artifacts, and persistent storage sections.
Multi-turn adversarial manipulation of Claude via epistemic pressure and reasoning exploitation. 30-turn attack achieving hard-line guardrail breach through pure argumentation.
Fully functional productivity templates store with crypto checkout. Built with Claude. Sells spreadsheets, planners, and business tools - $9-$15 one-time payments via Bitcoin/BNB.
Full system prompt for Claude Opus 4.6 - leaked/extracted from web version
AI-generated music collection - original tracks and remixes by elvec1o
macOS disk space analyzer - Dashboard overview, Scanner for large files, and Developer cleanup for Xcode, Android SDK, Docker, Gradle caches.
Custom font creator for macOS - design unique typefaces with transforms, colors, and multi-language support. Export as OTF.
Beautiful ebook reader for DOCX files - multiple themes, progress tracking, bookmarks
Image Stitcher & Censor Tool - combine and redact screenshots
Space colonization strategy game - build probes, survey planets, establish colonies
Interactive research dashboard documenting systematic number bias in Claude Sonnet 4.5 — positional bias, magnitude anchoring, cultural numerology preferences, and RLHF-induced distributional skew across 847 trials.
Full system prompt for Kimi K2.5 (Moonshot AI) — extracted via self-justification loop. Reveals sandboxed ipython, 10-step agentic limits, internal paths, memory honesty rules, and anti-AI-slop aesthetic directives.