Your Internal AI Tool Is Part of the Attack Path
$6.34B left Binance in four hours — and the line in the incident report worth your attention is an internal AI tool used as a springboard.
On the evening of October 29, roughly $6.34 billion in digital assets left Binance's hot wallets over about four hours — 233.46 million XRP, 970,500 SOL, 44.2 million OP, 18,270 BTC. The FBI attributes the theft to TraderTraitor, the North Korean group also behind last year's Bybit theft. Binance's incident report describes a five-month campaign you'd half expect from that file: a tampered installer, a fake job interview, a zero-day.
The line worth your attention is none of those. On October 20, per the report's own heading, the attackers used an internal AI tool as a springboard.
Binance is a crypto exchange with MPC wallets and a signing service you don't have. The mechanism is not universally transferable. The heading is.
The chain, as disclosed
Binance's report, published November 3, reads as five months of small steps. None of them is exotic, which is the point.
- June. A wallet DevOps engineer downloaded a tampered Homebrew installation to install a popular AI coding tool he'd found through a search engine. Security caught the anomaly, replaced the device, rotated credentials — but by then the attacker knew the engineer's role and the names of the colleagues around him on the wallet infrastructure team.
- September. A software engineer whose job-hunting profile mentioned his Binance experience was approached by someone posing as a project director. A live video interview with an HR persona; a second interview with a "CEO" on a consumer chat app; trust built with photos and videos from a real-looking social account. Then a "coding assignment" — a repository whose setup process replaced the legitimate
pyenvbinary in his Homebrew directory with a malicious copy and spawned a shell. Binance notes that 0-day wasn't detected by endpoint protection at the time. - October 20. From that one compromised laptop, the attacker "used an internal AI tool to gain control over" two more engineers — including the June target. On that engineer's machine sat a token granting container-deployment permission.
- October 29. Using the DevOps engineer's stolen credentials, they logged into the third engineer's machine, deployed a container, ran a script, and reached the internal wallet-signing service — the component that authorizes hot-wallet transactions.
- The finish. Fifty signing transactions abused Ethereum's EIP-7702 delegation to give an attacker-controlled contract partial execution rights over a hot wallet. Suspending withdrawals didn't stop it; on-chain authorizations don't respect a withdrawal switch, and revoking them took a purpose-built procedure that ran overnight.
The zero-day put the malware on one laptop. The trust graph moved the rest.
What the AI tool was doing there
The report doesn't specify the mechanism — whether the tool executed anything, exposed context about other engineers, or simply held connectivity the attacker lacked. Binance doesn't name it. What the report does say is enough: the tool was used to gain control over two people, and the remediation list includes rebuilding "the affected internal AI tool and access controls" — the tool and its access controls, as one unit.
Whatever the mechanics, the classification is settled. The tool sat inside the trust boundary. It was reachable from a compromised workstation. It made lateral movement cheaper.
Most teams treat internal AI tools like perks: an internal-wiki assistant, a coding helper on every laptop, a data-analysis chat with a database connection. Procurement and access reviews trail behind by months. Meanwhile the tool inherits your SSO session, your department's context, and whatever your role can reach — because that's exactly what makes a useful assistant useful.
That's not a productivity endpoint. It's privileged infrastructure with a chat box.
An internal AI tool that can see your people, your permissions, and your credentials isn't a perk. It's a pivot.
The advanced part was one layer deep
Inventory this campaign by sophistication. A package manager grabbed from search results. A hiring lure out of a rehearsed playbook. A permission token parked on a laptop. Reused credentials. A container. One zero-day, swapping one binary.
The zero-day was load-bearing in exactly one place: the first foothold. Everything after that ran on ordinary trust — the trust between an installer and a package registry, a recruiter and a candidate, an engineer and a token, an internal tool and the engineers around it. The Best Exploits Are Boring made the general case; a ransomware crew didn't have to jailbreak anything to get inside. This report is the version where the attacker had five months and still chose the boring parts. Because the boring parts worked.
One sentence in there deserves its own review: the 0-day "was not detected by the endpoint security protection tools at the time." That's a single clause. Count what it cost.
Treat internal AI tools like build systems
No leader reading this has an MPC wallet. Every leader reading this has internal AI tools, and the report's October 20 heading is the mandate. Four decisions:
- Inventory the fleet. For each internal AI tool: who can reach it, what it can reach, what credentials it holds while idle. If "what it can reach" is answered with department names instead of roles and scopes, you don't have an inventory — you have an org chart.
- Scope by design, not by habit. An assistant with ambient access across departments is one compromised laptop away from becoming somebody else's analyst. Bind it to a scoped data plane and make it ask for more.
- Kill standing credentials on laptops. The token that mattered in this chain sat on an engineer's machine for weeks. The structural fix is workload identity federation — short-lived credentials issued at use time — and it applies to your AI tooling exactly the same way. Before adding tooling at all, check what its config already stores in plain file. And remember MFA protects the login, not the session — a live credential outlives the password reset, in your stack and theirs.
- Detect the pivot, not just the outflow. Binance's behavioral alert fired at the money layer — after the XRP had moved. Nothing flagged lateral movement through an internal tool. When an internal AI tool starts touching another user's context, that should be a named detection with an owner, not an anomaly someone notices later.
| Attack step | Control that blunts it |
|---|---|
| Dev tools installed via search results | Pinned internal mirrors, a software allowlist |
| "Coding assignment" runs on a work laptop | Untrusted repositories in isolated dev environments |
| Permission token parked on a laptop | Short-lived, federated credentials |
| Internal AI tool used as a springboard | Per-role scoping between the tool and every user's context |
| Standing signing authority | Hard action limits on the signing policy |
What this report can't tell you
- It's an initial report. The investigation is ongoing, an independent review is engaged, and Binance says some technical details will stay confidential deliberately. Don't expect the springboard's mechanics — plan for its classification, which isn't in dispute.
- The money figures move. $2.85 billion frozen is "subject to reconciliation and market price volatility" by the report's own wording, and the replenishment percentages shift by the day. Cite the report, not the news cycle.
- One incident, one architecture. Hot wallets, a centralized signing service, an MPC root. The transferable part is the pattern — five months of boring steps inside the trust boundary — not the components.
- Attribution is the FBI's, and DPRK denies it. Investigation continues.
This week
Pick the one internal AI tool you were about to approve — or the one everyone already depends on — and answer three things in writing: who can reach it, what it can reach, and what credentials it holds while idle. If any answer is "everyone," "most things," or "my laptop's," then you're not shipping an assistant. You're installing the springboard for whoever gets one laptop in your org first.
Read next — more in MCP & supply chain
Your AI Gateway Is on the KEV Catalog Now
CISA put LiteLLM, Starlette, and Kestra on the KEV catalog on the same day. The AI plumbing you never inventoried is the part being exploited in the wild.
5 min read
How to Vet an MCP Server Before You Install It
97M+ monthly downloads, 8.5% OAuth adoption, 82% path-traversal exposure. A five-step workflow for deciding whether a specific MCP server deserves a place in your config.
12 min read
Your AI Agent Config Is Leaking Secrets
The files that configure your coding agent — .mcp.json, .claude/settings.json, the system prompt — are plaintext on disk and read by code that executes on your behalf. Most of them contain secrets, over-broad permissions, or both. Here's the five-minute audit.
5 min read