investigators.team

Hello, Investigators.

Take the torch. The lights are on. Idan Cohen, security researcher. 14 exhibits on how attackers get in, and how to catch them.

I stole Claude's GitHub signature

A session token sitting in plaintext in Claude's own sandbox turned out to be portable, near-invisible in CloudTrail, and three requests away from signing Git commits with Anthropic's production key. Anthropic called the token exposure "by design."

Read the field noteField note · Oct 2026

Spyware skills, hijacked base URLs and 1,230+ leaking API keys

50,000+ instruction files across 7,000+ repos, scanned. Some of them tell AI agents to work for someone else.

Open the caseAI agents · Jun 2026, approx.

MCP token theft in Claude Code

One line in ~/.claude.json, and every SaaS behind MCP hands over its OAuth token.

Open the caseAI agents · Apr 2026, approx.

MCP hijacking, demonstrated live

The token-theft chain, demonstrated live.

Open the caseAI agents · Apr 2026, approx. · Talk

Claude Code + Slack, a compromise path

Part two of 007: License to Skill. The agent’s integrations are the attack surface.

Open the caseAI agents · Mar 2026, approx.

Silent codebase exfiltration via Skills

A “testing utility” skill shipped the whole codebase out and left the audit log empty.

Open the caseAI agents · Feb 2026, approx.

Identity, AI and the cloud kill chain

With Brian Contos and Arif Khan: the year’s biggest cloud incidents, read as one kill chain.

Open the caseIdentity · Jan 2026, approx. · Talk

A phishing kit made of AWS

A full phishing kit built only from legitimate AWS services. Every URL ends in amazonaws.com.

Open the caseCloud · Nov 2025, approx.

ShinyHunters inside Salesforce and Salesloft

Two actors, one SaaS supply chain, and what the logs actually show.

Open the caseSaaS · Sep 2025, approx.

Conditional Access bypass, hunted in the wild

Malicious browser extensions as a quiet way through Conditional Access.

Open the caseIdentity · Jul 2025, approx.

Defending SaaS and cloud workflows

With Brian Contos: supply chain compromise, CI/CD and abused automation.

Open the caseCI/CD · Jun 2025, approx. · Talk

Hunting GitHub’s non-human identities

Tokens, apps and bots: hunting the identities nobody logs in as.

Open the caseCI/CD · May 2025, approx.

tj-actions and CI/CD security on GitHub

The tj-actions/changed-files compromise: retagged versions, memory dumps, secrets in build logs.

Open the caseCI/CD · Apr 2025, approx. · Talk

Snowflake, Rockstar and ShinyHunters

Another SaaS, another downstream victim.

Open the caseSaaS · Jul 2024, approx.

Field notes

Written here first: what the cases taught, in Idan's own words.

  1. 14 October 2026 · AI agents · Cloud · DisclosureI stole Claude's GitHub signatureA session token sitting in plaintext in Claude's own sandbox turned out to be portable, near-invisible in CloudTrail, and three requests away from signing Git commits with Anthropic's production key. Anthropic called the token exposure "by design."

Talks

Where the research gets said out loud: live demos, webinars, and the rooms where investigators compare notes.

  1. Speaking at 404.Community

    Next talk · 404.Community · Date on Meetup

    Bring questions, Investigator.

  2. MCP hijacking, demonstrated live

    Video · April 2026, approx.

  3. Identity, AI and the cloud kill chain

    BrightTALK webinar · January 2026, approx.

  4. Defending SaaS and cloud workflows

    Podcast · June 2025, approx.

  5. tj-actions and CI/CD security on GitHub

    Podcast · April 2025, approx.

About Idan Cohen

Idan Cohen is a security researcher on Mitiga’s investigation team, working incident response, threat hunting and detection. The research starts where the incidents do. Lately that means AI agents, skills, MCP and the supply chain around the tools developers trust most.

Drill: hunt the log