In March 2023, attackers walked into Okta's support system using a stolen employee credential: no malware, no exploit, no phishing email that morning. They used a session that already existed.
Six months later, a similar pattern hit MGM Resorts: a ten minute phone call to a help desk, a reset password, and a ransomware crew was inside a casino operator's network without touching a single firewall rule. Neither incident involved breaking through a perimeter. Both involved becoming someone who was already allowed in.
This is not an edge case anymore. It is the default. For most of the last two decades, security budgets were built around the assumption that the network edge was the thing worth defending: firewalls, VPNs, intrusion detection, segmented subnets. That assumption made sense when infrastructure lived in one data center with one entry point. It makes much less sense when a single mid size company now runs workloads across AWS, a neocloud GPU provider, a managed database vendor, and a handful of SaaS platforms, each with its own identity system and its own set of credentials, tokens, and service accounts. The fence has no edges left to guard. What is left is identity: the set of credentials that decide who, or what, is allowed to act.
The traditional security model assumed a trusted inside and an untrusted outside, with the network boundary as the control point. Cloud broke that model in three specific ways that matter for infrastructure buyers, not just security teams.
First, workloads no longer sit behind a single boundary. A training job might read data from an object storage bucket, call a GPU cluster at a separate provider, and write results back through a third vendor's API, none of which share a network perimeter. Second, most access into that chain now happens over the public internet by design, authenticated by a credential rather than by network location. Third, the number of things that need credentials has grown faster than the number of people. Every microservice, every CI/CD pipeline, every AI agent calling a model endpoint needs its own identity, and each one is a potential entry point that a firewall was never built to see.
The result is that the question security teams ask has changed. It used to be: can an attacker get through the wall? Now it is: can an attacker convince the system they are already supposed to be inside it? That second question is far harder to answer with network tooling, because a stolen credential looks, to most logging systems, exactly like legitimate access.
Credential based intrusion has a structural advantage over network intrusion: it does not need to defeat anything. A network attack has to find an unpatched service, an exposed port, or a vulnerability worth weaponizing, all of which leave forensic evidence and can be blocked by a patch. A credential attack simply reuses a password, a session token, or an API key that was never revoked. There is no exploit to detect because nothing is being exploited. The system is functioning exactly as designed, for the wrong person.
Modern phishing kits do not try to guess a password anymore. They proxy the real login page in real time, capture the session token issued after multi factor authentication succeeds, and replay that token before it expires. The victim sees a normal login. The attacker walks away with a session that is already fully authenticated, no password cracking and no MFA bypass required in the traditional sense, because the MFA step already happened, just for the attacker's benefit instead of the user's.
Because credential based access blends into normal traffic patterns, it tends to go undetected far longer than malware or exploit based intrusions. Attackers using valid accounts can browse quietly, escalate privileges gradually, and exfiltrate data in patterns that resemble routine business activity, which is exactly why security teams increasingly describe identity, not the network, as the front line.
| Attribute | Network Intrusion | Credential Based Intrusion |
|---|---|---|
| Primary defense | Firewalls, patching, segmentation | Credential hygiene, MFA, least privilege |
| Typical detection method | Intrusion detection, traffic anomalies | Behavioral analytics, spend anomalies |
| Leaves a forensic trace of | An exploited vulnerability or malware artifact | Nothing: it looks like normal access |
| Common first sign | An alert from IDS or endpoint tooling | An unexpected cloud bill or unusual data egress |
The identity conversation usually centers on humans: employees, contractors, the people who log into a VPN or an SSO portal. That framing is already out of date. In modern cloud and AI infrastructure, the overwhelming majority of credentials in circulation belong to things, not people: API keys for a data pipeline, service accounts for a Kubernetes cluster, OAuth tokens issued to a third party integration, and the access tokens an AI agent uses to call a model endpoint or query a storage bucket on a human's behalf.
This matters specifically for AI infrastructure teams because agentic workloads multiply machine identities faster than almost anything else in the stack. A single production AI agent might hold credentials to a vector database, a model provider, an object storage account, and an internal API, each one a standing grant of access that persists long after the task that created it is finished. Unlike a human employee, a stale service account does not get flagged in an offboarding checklist. It just sits there, valid, until someone finds it, and increasingly, that someone is an attacker rather than an auditor.
The practical risk is not abstract. A leaked API key for a GPU provider or a cloud storage account is not just a data exposure question, it is a resource theft question: someone else can spin up compute or move data using an account that is not theirs, and the bill lands on the legitimate owner before anyone notices the anomaly. Teams that never audit for orphaned service accounts often only catch the problem the same way they catch a cloud misconfiguration: after the invoice arrives.
Security teams measure credential compromise in breach costs. Infrastructure and FinOps teams should measure it in a second currency: the compute and egress bill an attacker can generate before detection. Both numbers matter, and they compound each other.
A compromised cloud credential with permission to launch instances is, in practical terms, a blank check against a company's GPU spend. Cryptomining and unauthorized model training runs on stolen accounts are common enough that several major cloud providers now run automated anomaly detection specifically for sudden compute spikes tied to a single credential. The same logic applies to data. An attacker who authenticates as a legitimate service does not need to defeat encryption or storage security, they simply request the data through the front door and pay whatever egress fee the provider charges, except the bill goes to the victim, not the attacker, turning a security incident directly into a cost anomaly on next month's invoice.
This is also where storage architecture and identity security intersect more than most teams expect. Zero egress providers like Backblaze B2 do not just reduce the cost of routine multi cloud data movement, they also cap the financial damage an attacker can do by exfiltrating data through a compromised credential, since there is no per gigabyte toll for the attacker to run up on the way out. That is a genuine architectural argument for egress free storage in an identity first threat model, not just a cost optimization one.
None of this means network security is worthless, only that it is no longer sufficient on its own. Infrastructure teams that are serious about the shift toward identity based attacks tend to converge on the same handful of practices, and most of them overlap heavily with the discipline behind a zero trust architecture.
The network perimeter did not disappear, it just stopped being the thing worth defending most. In a multi cloud, multi vendor, agent heavy infrastructure stack, identity is the perimeter now, and it is far more porous than the firewall it replaced.
Compare AWS, Google Cloud, Azure, and alternatives like Backblaze B2 Discover how much you could save in seconds
