AI Security Threats Go Live: Infostealers Are Now Hijacking Claude Sessions

ai security threats — featured image

On August 30, Anthropic quietly warned a subset of Claude users that infostealer malware on their machines had lifted active login sessions and handed attackers a free ride on their accounts — burning through paid usage, reading conversation history, and in some cases reaching the API. Bleeping Computer broke the story here, and it is the cleanest proof-of-concept yet that ai security threats are not a slide-deck abstraction. They are cookies stolen out of Chrome memory last Tuesday.

If you are trying to figure out whether AI security is real work or vendor theater, this incident is your answer. The category of ai security threats that matters most right now is not model jailbreaks or prompt injection demos — it is the boring, mechanical stuff: session tokens, endpoint compromise, and account takeover applied to a new class of high-value target.

The Anthropic Wake-Up Call: When AI Usage Becomes Currency

The mechanic is simple and it is why this matters. An infostealer landed on a user’s PC — probably through a cracked installer, a malicious npm package, a fake browser update, the usual delivery chain. Once resident, it did what infostealers do: it scraped browser storage. But instead of just grabbing saved passwords and card numbers, it walked away with the session cookies for claude.ai.

Those cookies are bearer tokens. Present the cookie, you are the user. No password prompt, no MFA challenge, no email alert. The attacker imports the cookie into their own browser and Claude opens up as if they had logged in themselves. They then do the economically rational thing: they run the account into the ground. Long-context prompts, expensive model calls, batch jobs against the API if a key is exposed in the workspace. The usage meter spins, and the legitimate user gets the bill or the rate-limit wall.

Anthropic is not the villain of the story — they detected the abuse pattern and notified affected accounts. But the story is the pattern itself. Session cookies from AI platforms are now on the shopping list next to bank logins and crypto wallets, because they convert directly to compute, which converts directly to money. That is what ai security threats look like once the economics catch up.

How Infostealer Malware Actually Works (What Beginners Miss)

Most people entering the field assume malware steals credentials by logging keystrokes. That was true fifteen years ago. Modern infostealers — Lumma, RedLine, Vidar, StealC, and the endless rebrands — barely touch the keyboard. They go straight for what is already unlocked in memory, and that shift is what makes today’s ai security threats so hard for newcomers to reason about.

Here is the actual workflow on a Chromium-based browser, and it is worth understanding at this level of detail if you want to reason about defenses:

  • Chrome stores cookies in an SQLite database at %LocalAppData%\Google\Chrome\User Data\Default\Network\Cookies.
  • Sensitive values are encrypted with a key stored in Local State, itself wrapped by Windows DPAPI and tied to the logged-in user.
  • Because the malware runs as that user, it can call DPAPI and decrypt the key. No privilege escalation required. This is the design.
  • Newer builds bypass Chrome’s App-Bound Encryption by injecting into the browser process or abusing the elevation service, so even the hardened cookie store is not a durable defense.
  • The stolen cookies, autofill data, extension state, and any active tokens get bundled into a ZIP and posted to a Telegram bot or a C2 panel within seconds.

The cookie for claude.ai — or ChatGPT, or Gemini, or your Copilot tenant — sits in that same jar. If it has a long lifetime and no device-binding, it is portable. Portable means saleable. There are already Telegram markets moving “AI logs” as a distinct category. The Anthropic advisory is the moment that category went public and the moment ai security threats stopped being a category anyone could dismiss as hypothetical.

Reading Anthropic’s own threat intelligence work on account misuse alongside this incident is worth twenty minutes — it shows how the same team defends the platform from both outside abuse and inside compromise.

Why AI Platforms Are Session-Hijacking Gold Mines

Session hijacking against a bank is loud. The bank has fraud models trained on decades of behavior, device fingerprinting, step-up auth on any meaningful action, and a legal obligation to make the victim whole. Draining a bank account triggers alarms the moment the wire leaves.

AI platforms are the opposite environment, and that asymmetry is why ai security threats have moved up the attacker priority list so fast:

  • Long session lifetimes. Nobody wants to re-auth into their assistant every four hours. Cookies last days to weeks. That is a huge window for a stolen token.
  • High-value, fungible usage. A hijacked session is a compute faucet. Attackers resell it, use it to generate spam or scam content at scale, or wire it into their own downstream automation.
  • API keys in reach. A logged-in session often has one click to the console where an API key can be minted or exfiltrated — and API keys usually outlive session cookies by an order of magnitude.
  • Weak behavioral baselines. The platforms are new. “This user suddenly ran 400 long-context prompts at 3am from a Moscow IP” should be a red flag, but the detection stacks are still being built. Compare that to a bank’s fraud system, honed over thirty years.
  • Business context in the transcript. The stolen session is not just compute — it is a searchable archive of the victim’s work. Source code, contracts, internal memos, credentials pasted for “help me debug this.” That is arguably worth more than the usage.

This is the same pattern we called out in our recent write-up on the AI-pentesting career angle emerging from Anthropic’s Fable and Mythos work, and it echoes the injection-surface analysis in our piece on brain-computer-interface prompt injection as an emerging class of ai security threats: attackers are treating AI platforms as first-class targets, and the defensive side is playing catch-up on fundamentals that other verticals figured out a decade ago.

What This Means for Security Careers: The AI Attack Surface Is Here

Here is the part I care about, because it is the part most beginner-oriented content on ai security threats skips. You do not need to be an ML researcher to work on this problem. You need to be a competent security engineer who understands where AI platforms fit into the classical attack surface. The Claude session-hijacking story is entirely a classical problem: endpoint compromise, browser token storage, session management, anomaly detection. Every one of those is a lab-testable skill.

Concretely, the work of defending against ai security threats at this layer looks like this:

  • Threat-modeling an AI integration. Where does the token live? What is its lifetime? What can an attacker do with it if it leaks? Where is the human-in-the-loop, and where is it not?
  • Testing session storage. Pull cookies from a test browser profile. Replay them from a different machine. Prove or disprove that the platform binds sessions to device fingerprints. Most do not.
  • Understanding token lifecycles. OAuth refresh tokens, API keys, session cookies, service-account JWTs — how they interact, how they get rotated, how they get stolen. This is table-stakes and it is where most “AI security” job descriptions actually live.
  • Mapping the endpoint chain. Infostealer defense is endpoint defense. EDR tuning, browser hardening, DPAPI hygiene, extension inventory. The AI angle is a delivery target; the discipline is unchanged.
  • Extending the muscle to adjacent surfaces. The same token-and-session reasoning applies to IoT command channels and vehicle backends — if you want the embedded angle, our IoT hacking course and car hacking course both drill the same session-and-token muscle on hardware targets, which is where a lot of AI-adjacent work is heading.

The reason the Anthropic advisory is worth writing about is not that Claude is broken. It is not. The reason is that it collapses the distance between “AI security” as a buzzword and ai security threats as concrete engineering problems you can start solving this week with skills that transfer from web app pentesting, endpoint work, and identity engineering. If you can reason about a stolen cookie, you can reason about this. If you cannot yet, that is exactly the gap to close.

The full LufSec course catalog is organized around that transferable core — offensive and defensive fundamentals first, then the vertical (web, IoT, automotive, and increasingly AI-integrated systems) on top. That order matters, because the vertical churns and the fundamentals do not.

Start Where the Work Actually Is

If this piece landed and you want a structured way in, the LufSec Cybersecurity Career Guide is free and it is built for exactly the reader who just followed the infostealer walkthrough above and thought “okay, I could learn this.” It maps the paths into the field — analyst, pentester, appsec, cloud, and the newer AI-adjacent roles that are being defined right now by incidents like this one — against the skills you already have, tells you honestly which certs are worth your time and which are not, and points you at the hands-on labs that get you interview-ready rather than just quiz-ready. It is written for the person entering or repositioning inside cybersecurity, not for someone who already has a decade in. Grab it, work through it, and use the Anthropic story as your worked example of why the fundamentals in that guide are not optional. The ai security threats are already here. The people who can parse them are the ones the market is hiring.