All work.

Original Beyond the Model episode cover featuring Shmulik Cohen

מעבר למודל: מי בוחר את המילה הבאה?

Guest conversation with Dekel Asaf and Shiron Babi about sampling, softmax, temperature, Top-K, Top-P, and Min-P. Episode 4, in Hebrew.

AI / SamplingOpen original

Evil Qwen’s Jewish Cousins

A playful exploration of steering a small language model through its internal activations.

AI / InterpretabilityOpen original

Jewish Qwen

Experiments in steering Qwen3.5-4B with neurons, activation directions, and optimized vectors.

AI / InterpretabilityTake a look

MCP TypeScript SDK (@modelcontextprotocol/hono): HTTP body / JSON-RPC batch denial of service

Public maintainer advisory. Reporter credit may be shared with other researchers.

OpenMage: restricted-admin authorization bypass

Authorization gaps in sales and order operations. The advisory narrative credits shmulc8 for the reported cases.

Twenty: pending-invite takeover

An unauthenticated signup path exposed a pending invitation token. Public advisory credits shmulc8.

Original Darkroom article cover: a photograph developing in a red-lit darkroom

Darkroom: malware hidden in a photo

Tracing npm dotenv look-alikes that concealed a PowerShell payload inside a JPEG.

Security / Supply chainTake a look

The Paste That Runs the Fleet: Inside CheatSheet, a 1,000-Repo Malware Network

Investigating a malware distribution network spread across GitHub and GitLab.

Security / AIOpen original

1 | BERT

How BERT changed natural language processing and contextual understanding. Prompt and a Half, with Roy Benjamin and Shmulik Cohen.

AI / BERTOpen original

2 | I Don't Want to Become a Meat Proxy

Using AI while preserving independent judgment, understanding, and practical value. Prompt and a Half, with Roy Benjamin and Shmulik Cohen.

3 | סוכנים אישיים והמסע של מודלי השפה

Personal AI agents and the evolution of language models. Prompt and a Half, with Roy Benjamin and Shmulik Cohen.

AI / AgentsOpen original

Original Apiiro cover for the Argo CD MCP vulnerability article

Argo CD MCP: unauthenticated cluster access

An investigation into the authentication boundary between an AI tool and the infrastructure it controls.

Security / MCPTake a look

BleepingComputer: FakeGit

Independent coverage of Apiiro’s FakeGit investigation, linking the original report. Coverage credits Apiiro’s research team.

Security / Supply chainRead coverage

Original Apiiro cover for Never Deleted, Only Re-Pointed

Never Deleted, Only Re-Pointed: Inside the 17,600-Repo FakeGit Fleet That Re-Arms Overnight

Following README link changes across the FakeGit software-supply-chain campaign.

Security / Supply chainTake a look

Digital Craft Workshop: Show Me Your Bench

An interview about AI Superhero, Substack Network, and my research and writing workflow.

AI / EngineeringRead coverage

CyberKart

A cyber-themed kart racer with Guardian, parody characters, and a playable browser demo.

Games / ExperimentsTake a look

I Don’t Want to Become a Meat Proxy

A personal essay about keeping your judgment and curiosity while building with AI.

MCP Python SDK: request-body exhaustion

CVE-2026-109527. Unbounded HTTP and OAuth body reads. Consolidated advisory with shared reporter credit.

You Can Call Me Shmulik Linux

The story behind contributing boundary checks to upstream Linux Wi-Fi drivers.

OpenMetadata: SAML login redirect leaks a JWT

An open redirect in the SAML login flow could expose a JWT to an external host. The public advisory credits my work account, shmulik-apiiro.

A fruit-fly brain learns to blow a shofar.

A connectome-inspired simulation, a browser, and an unusual musical goal.

AI / ExperimentsOpen original

FlexGet: archive extraction outside the destination

A malicious archive processed by the decompress plugin could escape the extraction directory and lead to code execution. The maintainer advisory credits shmulc8.

Bazarr: Windows PWA path traversal

An unauthenticated arbitrary-file-read issue on Windows. Public advisory credits shmulc8.

Original cover artwork for That Escalated Quickly

That Escalated Quickly: 7 Permissions That Are Bigger Than They Look

A closer look at permissions and the authority they grant to agents.

MCP Java SDK: request-body exhaustion

CVE-2026-108395. Unbounded body reads in HTTP transports. Consolidated advisory with shared reporter credit.

How Many Exploits Does It Take to Turn an Egg into a Legendary Roc?

An agent experiment involving a game, its rules, and some unexpected behavior.

Full first page of Exact Metric Dimensions and Improved Upper Bounds for Hypercubes by Shmulik Cohen

Hypercube metric dimension

Computational experiments and machine-checkable certificates for hypercube metric dimension.

Math / Open sourceTake a look

Can GPT Still Be Fooled?

A prompt-injection experiment with published fixtures, scoring, and results.

Can GPT still be fooled? The experiment.

The code, prompt fixtures, scoring details, and published data behind a prompt-injection comparison.

AI / SecurityOpen original

Finding 6 CVEs in Open-Source AI Tools with Apiiro AI-SAST

Collaborative research into vulnerabilities in open-source AI tools. The original article distinguishes each researcher’s findings.

Security / AIOpen original

Original cover artwork for Building My-Little-Substack

Building My-Little-Substack: An On-Device AI Feed Lens

Building a browser tool that uses on-device AI to help make sense of a feed.

AI / ResearchOpen original

לפצל את השברים: בינה מלאכותית ומגילות קומראן

Coauthored with Mike Erlihson. A Hebrew review of NLP methods for grouping fragmented Qumran scrolls, combining semantic representations with statistical language features.

HebrewAI-curious

n8n: SQL injection in MySQL node

CVE-2026-59257. Public maintainer advisory. Reporter credit may be shared with other researchers.

Hebrew Codenames

A semantic-alignment experiment using Hebrew word embeddings, model comparisons, and an interactive map.

AI / EmbeddingsTake a look

Captain Obvious

A deterministic scanner for tests that can never fail, using type-checker and program-structure analysis.

Open source / TestingTake a look

EnglishBuilders · Security practitioners

Packing 101: from Windows 2009 to Apple Silicon

Five runnable lessons rebuilding a classic Digital Whisper executable-packing tutorial for macOS and arm64, with an animated unpacking view.

Security / Reverse engineeringTake a look

Original cover artwork for Two Open Models vs the Security Benchmark

Two Open Models vs the Security Benchmark That Humbles the Frontier

Comparing open models on a software-security benchmark.

Secure by Design, Broken by a Filename

How inbound and outbound attachment filenames crossed NanoClaw’s host-side file boundaries, and how the maintainer fixed both paths.

NanoClaw: path traversal outside the sandbox

Two filename-validation gaps in host-side attachment handling: an inbound write and an outbound read. The public writeup links both maintainer fixes.

Security / AI agentsTake a look

Text Is a Terrible Way to Store Text

An exploration of how we represent and work with text.

Original cover artwork for Project Hail Mary

Project Hail Mary: Building the Ultimate Offline AI Coding Rig

Exploring the engineering choices behind an offline AI coding environment.

AI / ConversationOpen original

Media2026

Original Data & AI podcast artwork supplied by the author

DATA & AI: OpenClaw and the agent toolbox

A guest conversation about AI agents, the tools around them, and practical engineering choices.

Hebrew onlyAI-curious

Every CVE Unlocks Another One

Lessons and connections from vulnerability research.

Codex Pet Platformer

A customizable side-scrolling platformer built from a Codex pet sprite, packaged as an agent skill and deployable static site.

AI / ExperimentsTake a look

Talk2026

Codex Pets: playful agents, real possibilities

My talk at the first Codex Community Meetup at Wix, covering Codex Pets, Guardian, and the possibilities opened by a shared protocol.

AI / ExperimentsTake a look

WebSocket vs SSE: visualized

An animated side-by-side demo that makes streaming, reconnects, and network events visible in real time.

AI / NetworkingTake a look

Parse Server: Unauthenticated regular-expression DoS

CVE-2026-47138. Public maintainer advisory. Reporter credit may be shared with other researchers.

The Hacker News: PraisonAI authentication bypass

Reporting on CVE-2026-44338 and post-disclosure scanning, explicitly crediting Shmulik for discovering and reporting the vulnerability.

Security / AI agentsRead coverage

Sysdig: PraisonAI after disclosure

Independent threat research on scanning observed after disclosure of CVE-2026-44338, the PraisonAI authentication flaw reported by Shmulik.

Security / AI agentsRead coverage

PraisonAI: Injection via SQL/CQL collection names

CVE-2026-44337. Public maintainer advisory. Reporter credit may be shared with other researchers.

PraisonAI: Missing API authentication / agents workflow access

CVE-2026-44338. Public maintainer advisory. Reporter credit may be shared with other researchers.

PraisonAI: Unsafe callable resolution / tool execution

CVE-2026-44339. Public maintainer advisory. Reporter credit may be shared with other researchers.

AI / ResearchOpen original

האם AI יכול להחליף מפתחים במחקרי הנדסת תוכנה? הפתרון לפער בין האקדמיה לתעשייה

Coauthored with Mike Erlihson. A Hebrew review asking whether AI can stand in for developers in software-engineering user studies, and what that means for reliable evidence.

HebrewAI-curious

open-webSearch: SSRF / private IP filter bypass

CVE-2026-42260. Public maintainer advisory. Reporter credit may be shared with other researchers.

pyload-ng: Authorization flaw allowing TLS verification to be turned off

CVE-2026-42312. Public maintainer advisory. Reporter credit may be shared with other researchers.

pyload-ng: SSRF / proxy configuration

CVE-2026-42313. Public maintainer advisory. Reporter credit may be shared with other researchers.

HebrewAI-curious · Builders

Geektime: Israeli Claude Code

A profile of my Israeli output style for Claude Code, with an explanation of how the configuration works.

AI / ToolsRead coverage

Israeli Claude Code

An output-style configuration that brings Israeli expressions and a direct tone to Claude Code responses.

AI / ToolsTake a look

How I Found My First Two SSRF CVEs

A first-person account of researching server-side request forgery.

Agent Skills Network

An interactive 3D map of agent skills, grouped by semantic similarity.

AI / VisualizationTake a look

Mike’s Latent Space

An interactive semantic map of Mike Erlihson’s paper reviews.

AI / VisualizationOpen original

AI / ResearchOpen original

המהפכה של אנטרופיק במבחן - האם סקילים באמת עוזרים לסוכנים שלכם?

Coauthored with Mike Erlihson. The Hebrew review of SkillsBench: when agent skills help, hurt, or add noise.

HebrewAI-curious

Do Skills Actually Help Your Agents?

Coauthored with Mike Erlihson for Mathy AI. A review of SkillsBench and when agent skills help, hurt, or add noise.

AI / Paper reviewOpen original

EnglishBuilders · Security practitioners

SiliconANGLE: Apiiro CLI

Coverage of Apiiro’s AI-native security CLI, quoting Shmulik on building security tools that coding agents can use. The product is Apiiro team work.

AI / Developer toolsRead coverage

AlloyEval

A course project replicating and extending an evaluation of LLM-generated Alloy specifications.

AI / Formal methodsTake a look

From Idea to Product with Claude Code

The build story behind Substack Network, a Chrome extension for visualizing recommendation connections.

AI / ToolsOpen original

Give Your Agent a Breakpoint

DAP-based debugging tools for AI coding agents: breakpoints, runtime state, and practical limitations.

dapi: debug with your agent

An agent skill and DAP-based CLI for breakpoints, runtime state, and attach-to-process debugging across common languages.

AI / DebuggingTake a look

Stop “Vibe Merging”

AI code is exploding, but "Vibe Merging" is breaking production. A deep dive into the first neutral benchmark for Code Review agents: CodeRabbit, Cursor, Baz.

Original cover artwork for From World Cup Statistics to Business Data

From World Cup Statistics to Business Data: AI21’s New RAG

Coauthored with Mike Erlihson for Mathy AI. A hands-on review of Structured RAG for aggregative questions, with a reconstructed Colab implementation.

AI / RAGOpen original

Demystifying Coding Agents

From LLMs to MCP and Skills: Understand the simple concepts powering Cursor, Windsurf, Copilot and Claude Code.

50 Shades of BERT

Discover why BERT and ModernBERT remain essential in 2026. Learn how encoders beat generative LLMs in latency, cost, and precision for production NLP tasks.