https://samwallace.dev/ Application security, penetration testing, and vulnerability research, plus interactive walkthroughs of AI agent security, OAuth, and MCP. en-us Sam Wallace Sam Wallace Sun, 28 Jun 2026 12:00:00 GMT scripts/generate-rss.mjs https://samwallace.dev/mcp-gateway https://samwallace.dev/mcp-gateway Sun, 28 Jun 2026 12:00:00 GMT Interactive · Architecture Should an MCP server proxy your existing APIs, or go straight to the database? The MCP governance layer, with code: wrapping existing APIs, the catch with direct-to-DB MCPs, and how it scales across enterprise teams. https://samwallace.dev/oauth https://samwallace.dev/oauth Sat, 27 Jun 2026 12:00:00 GMT Interactive · Identity The friendly overview of OAuth 2.0 and OpenID Connect: the four roles, how delegated access works (code, then token, then API call), and where identity fits. The on-ramp to the PKCE and token-exchange deep dives. https://samwallace.dev/pipeline https://samwallace.dev/pipeline Sat, 27 Jun 2026 12:00:00 GMT Interactive · Threat model The AI agent attack surface, end to end. Toggle threats and defenses across the whole pipeline (RAG injection, the lethal trifecta, tool poisoning, excessive agency, output handling, memory poisoning, guardrail bypass) and watch the blast radius change. https://samwallace.dev/jailbreak https://samwallace.dev/jailbreak Sat, 27 Jun 2026 12:00:00 GMT Interactive · AI security Why jailbreaks and system-prompt extraction work: your input reshapes the model's next-token probabilities, so a refusal is only ever the most likely response, never a guaranteed one. Stack techniques and watch the meter flip. https://samwallace.dev/sandbox https://samwallace.dev/sandbox Sat, 27 Jun 2026 12:00:00 GMT Interactive · Isolation How scoping an agent's executor, not just its token, stops a hijacked tool call. Watch the same confused-deputy attack hit a sandbox boundary, then sail through with ambient root. https://samwallace.dev/toolcall https://samwallace.dev/toolcall Sat, 27 Jun 2026 12:00:00 GMT Interactive · Function calling What "function calling" actually does: the model proposes a structured call, your runtime validates it against a JSON Schema and decides whether to run it. Includes parallel calls and the security boundary. https://samwallace.dev/context https://samwallace.dev/context Sat, 27 Jun 2026 12:00:00 GMT Interactive · Tokens Where an LLM's token budget really goes (system prompt, tool defs, history, RAG), and how compaction reclaims space when the window fills up. https://samwallace.dev/mcpauth https://samwallace.dev/mcpauth Sat, 27 Jun 2026 12:00:00 GMT Interactive · OAuth 2.1 An interactive, click-through walkthrough of MCP authorization: how a client goes from a 401 to a scoped access token using OAuth 2.1, PKCE, dynamic client registration, and resource indicators (RFC 8707/9728). https://samwallace.dev/obo https://samwallace.dev/obo Mon, 01 Jun 2026 12:00:00 GMT Interactive · OAuth An interactive, click-through walkthrough of OAuth 2.0 Token Exchange (RFC 8693) on behalf of, for web apps, MCP servers, and agents. Watch which token goes where, step by step. https://samwallace.dev/research/stop-giving-your-ai-root https://samwallace.dev/research/stop-giving-your-ai-root Sat, 25 Apr 2026 12:00:00 GMT BSidesKC · Talk We hand AI agents API keys, OAuth tokens, and tool execution rights, then hope it works out. A defense in depth take on clawing control back: scoped cross-app access, input and output guardrails, and sandboxed execution. https://samwallace.dev/research/cve-2026-27699 https://samwallace.dev/research/cve-2026-27699 Mon, 23 Feb 2026 12:00:00 GMT Critical 9.1 · basic-ftp Path traversal (CWE-22) in basic-ftp's downloadToDir(). A malicious FTP server sends filenames with ../ and the client writes files outside the download directory. Critical, fixed in 5.2.0. https://samwallace.dev/research/ai-in-cybersecurity-how-to-be-a-10x-engineer https://samwallace.dev/research/ai-in-cybersecurity-how-to-be-a-10x-engineer Mon, 23 Oct 2023 12:00:00 GMT Talk · AI How to use AI as a 10x force multiplier. A hands-on tour of building LLM apps with LangChain, with examples anyone can copy. https://blog.dispareosecurity.com/can-ai-empowered-open-source-tools-compete-with-their-premium-counterparts-guest-blog-9683c3df71e5 https://blog.dispareosecurity.com/can-ai-empowered-open-source-tools-compete-with-their-premium-counterparts-guest-blog-9683c3df71e5 Tue, 28 Mar 2023 12:00:00 GMT AI Guest blog post on the Dispareo Security blog! https://samwallace.dev/research/dom-xss-skimming-credit-cards https://samwallace.dev/research/dom-xss-skimming-credit-cards Fri, 03 Mar 2023 12:00:00 GMT XSS Skimming credit cards by exploiting a DOM XSS vulnerability in a react application. This is for educational purposes only. https://samwallace.dev/research/harvesting-emails-with-expired-garmin-livetrack-sessions https://samwallace.dev/research/harvesting-emails-with-expired-garmin-livetrack-sessions Sat, 31 Dec 2022 12:00:00 GMT CVE-2022-46081 In Garmin Connect 4.61, terminating a LiveTrack session wouldn't prevent the LiveTrack API from continued exposure of private personal information. https://samwallace.dev/research/stored-xss-in-gms-1-0 https://samwallace.dev/research/stored-xss-in-gms-1-0 Mon, 12 Dec 2022 12:00:00 GMT CVE-2022-41358 Garage Management System utilizes client side validation to prevent XSS. Using burp, a request can be modified and replayed to the server bypassing this validation which creates an avenue for XSS. https://samwallace.dev/research/garmin-white-hat-hacker https://samwallace.dev/research/garmin-white-hat-hacker Sat, 01 Jan 2022 12:00:00 GMT Recognition Garmin credited me as a white hat for reporting a LiveTrack privacy bug (CVE-2022-46081).