The Boring Attacks Are Winning: Why Defenders Keep Losing to Trusted Tools
Leaked tokens, poisoned npm packages, and login replays are doing more damage than zero-days this quarter. Here is how to spot the pattern before it spots you.
The intrusions making payroll for criminals right now do not look like intrusions. They look like a Dependabot PR, a Slack OAuth grant, an Okta push you almost ignored. Attackers are not kicking doors. They are walking through the ones your team already propped open with trusted automation, and the telemetry to catch them often sits unread in a SIEM nobody tunes.
A quick tour of the last seven days makes the point. A fresh Linux rootkit family, tracked by Sysdig researchers, hides inside legitimate eBPF tooling that most EDRs treat as benign. A router zero-day in ASUS DSL-AC firmware (CVE-2025-59367) is being chained with default credentials to drop Mirai variants on small-business gateways. And npm picked up another wave of typosquats impersonating chalk and debug, this time exfiltrating GitHub tokens straight to a Telegram bot.
None of that is exotic. That is the point.
If you run a 5-to-50 person security team, the uncomfortable truth is that your alerting is probably tuned for the 2021 threat model: malware on endpoints, brute force on RDP, phishing with a payload. The 2025 model is quieter. It is a service account in Entra ID that suddenly enumerates mailboxes at 3 a.m. It is a CI runner pulling a package version that did not exist yesterday. It is a Citrix session token replayed from a residential proxy in the same ASN your sales team uses.
"The trust boundary moved from the perimeter to the supply chain and the identity provider, and most detection engineering has not caught up," said Katie Nickels, director of intelligence at Red Canary, in a recent briefing. Her team has been flagging a sharp rise in OAuth consent phishing against Microsoft 365 tenants, with Scattered Spider and a cluster Mandiant tracks as UNC6040 both leaning hard on the technique.
The AI angle is real but smaller than the headlines suggest. Prompt injection against agentic browsers is interesting research. Stolen Anthropic and OpenAI API keys getting resold on Russian Market for $4 to $40 a pop is the actual operational problem, because those keys are usually attached to a corporate billing account with no rate limit and no alerting.
So where do you spend the next afternoon? Probably not on a shiny new tool.
Five things worth doing this week:
- Pull a list of every OAuth app consented in your Microsoft 365 or Google Workspace tenant in the last 90 days. Revoke anything nobody can name.
- Rotate any AI provider API keys older than 60 days, and put a hard monthly spend cap on the billing account (not just the key).
- Add a detection for npm or PyPI installs of packages less than 14 days old on CI runners. Alert, do not block, for the first week.
- Re-check default credentials on every internet-facing router, firewall, and IPMI interface you own. Yes, including the one in the closet.
- Force a re-auth on any session token older than 24 hours for admin-tier accounts in your IdP.
None of that is glamorous. The attackers are counting on that.
