Who: Every organization that runs on cloud SaaS—startups, scale-ups, and enterprise stacks.
What: Nine seemingly innocuous default settings that, in August 2026, are still the fastest way for attackers to get persistent, legitimate access.
When/Where: August 2026 — across identity providers, email systems, collaboration suites, CRMs, CI/CD pipelines and AI connectors worldwide.
Why it matters: Attackers increasingly chain small, legitimate privileges—long-lived OAuth refresh tokens, perpetual guest accounts, non-expiring API keys—so breaches look like normal user behavior and dwell times extend. The defensive tools exist; most teams still run on convenience defaults.
Context: defaults are debt, and in 2026 that debt bites harder
We all learned the old checklist: enable SSO, turn on MFA, ship it. That stopped being enough by 2024. Since then, the industry has delivered better primitives: Continuous Access Evaluation (CAE), OAuth 2.1 best-practices, token attestation, ephemeral credentials, and Continuous Entitlement Management (CEM). IdPs including Microsoft Entra, Okta and Google have shipped richer JIT and entitlement controls. But adoption lags—and attackers live in the gaps.
In short: the tools improved; our defaults haven't. If your cloud stack still trusts "remember this device" cookies for admins, or allows org-wide OAuth consent by users, you have an open invitation for lateral access.
Updated checklist: the 9 settings attackers still love (and the August 2026 fixes)
1) “Remember this device” MFA sessions that last weeks
Why it still matters: refresh tokens and persistent session cookies are the new currency; they've replaced many password-based break-ins. Fix now: reduce privileged session lifetimes to 4–8 hours, require step-up auth for exports, API key creation and admin settings. Enable risk-based MFA with device posture checks and CAE where available so admins get strict controls while low-risk flows remain usable.
2) SSO exemptions and unmanaged break-glass accounts
Why it still matters: a single non-SSO local account is a permanent hole. Fix: inventory apps with local auth, disable local logins, or limit them to a single hardware-key-protected break-glass account stored in a vault with recorded access. Enforce event logging on any bypass and forward alerts to your SIEM/ITDR pipeline.
3) User-managed OAuth consent and broad third-party apps
Why it still matters: user-approved apps frequently request wide scopes (mailbox read/write, org-wide file access) that outlive churn. Fix: block user consent for write, offline, and org-scoped tokens. Implement admin-approval flows or a verified-publisher whitelist. Automate weekly scans for new org-scoped tokens and revoke anything not approved—use token-introspection APIs to trace active tokens back to an app.
4) "Anyone with the link" sharing and chat-ops leakage
Why it still matters: anonymous links leak via support tickets, chatbots, and AI assistants. Fix: default to "restricted/named users" for sensitive folders (security, finance, legal, M&A). Auto-expire external links after 7–30 days, apply pattern-scanning for SSNs/credentials on public links, and block public links for designated data classifications.
5) SCIM/JIT provisioning without enforced deprovisioning
Why it still matters: onboarding is automated; offboarding is not. Fix: require SCIM deprovisioning for core apps and integrate HR-system-triggered disable on termination dates. Pair SCIM with CEM so entitlements follow roles automatically. Add weekly reconciliation reports: "accounts in apps not present in IdP/user directory."
6) API keys and service tokens that never expire
Why it still matters: long-lived keys show up in repos, CI logs, and abandoned scripts. Fix: enforce scoped service accounts, require expiry (30–90 days) for keys, forbid personal tokens in production. Move CI/CD to short-lived, attested tokens (OIDC token exchange or workload identity federation) so you can rotate without breaking pipelines.
7) Overbroad admin roles granted for convenience
Why it still matters: coarse roles multiply blast radius. Fix: adopt a two-tier admin model—Day-to-Day Admin (limited tasks) and Super Admin (a handful of people, hardware-key MFA). Enforce time-bound elevation (1–24 hours) via IdP or PAM. Integrate JIT elevation with approvals and automated audit trails—no standing super-admin accounts.
8) Trusted domains and partner guest access that never expires
Why it still matters: partner ecosystems expand blast radius. Fix: remove permissive domain allowlists unless contractually required, place partners into segregated workspaces, and set guest access to auto-expire with re-approval. Apply DLP controls and session restrictions to guest accounts, and require contractual SLAs for data handling from vendors.
9) AI assistant connectors with permissive defaults
Why it still matters: AI connectors often request broad scopes for convenience—mailbox, drive, calendar—which creates exfiltration paths and training leakage. Fix immediately: treat AI connectors like payment processors—scope review, retention limits, "no model training" contractual clauses unless explicitly approved. Segregate AI connectors from legal/security/M&A folders and log every prompt and response for at least 90 days.
30-minute SaaS security reset (do this today)
- Shorten sessions for admin roles and high-risk apps (target 4–8 hours for admins).
- Disable user OAuth consent for write/offline/org-wide scopes; require admin approvals.
- Kill anonymous links org-wide; set automatic expirations and scan for sensitive patterns.
- Audit and time-box admin roles; convert super-admins to hardware-key only and enforce JIT elevation.
- List AI connectors and revoke any with org-wide write or train-on-data scopes; re-onboard only after review.
Repeat this at least monthly—ideally weekly if you have >500 seats or handle regulated data. Defaults drift; every contractor, tool, or integration adds entropy.
Impact and reactions
This matters for everyone. Small IT teams get hit hardest because they don't have staff to chase deprovisioning and token audits; CISOs worry about lateral movement rather than noisy ransomware. I spoke to two security leaders at SaaS firms in July 2026—both asked to remain anonymous. One told me: “We fixed MFA years ago and still got owned because a vendor's OAuth refresh token with mailbox access never expired.” The other said: “Our AI assistant leaked a contract because it was allowed to ‘help’ on a shared drive link.”
Regulators and boards are watching. The EU's cybersecurity frameworks and sectoral regulators continue to emphasize incident response and third-party risk; expect auditors to ask how you manage third-party OAuth tokens and AI connectors during 2026 audits.
What to watch next
- Wider adoption of token attestation and OIDC workload identity for CI/CD—reduce reliance on static keys.
- IdPs deepening CEM and JIT capabilities; expect richer APIs for entitlement lifecycle automation through 2026–27.
- More scrutiny and contractual standards for AI connectors—security teams should push for prompt logging, retention clauses, and “no-train” guarantees.
The takeaway
SaaS security is maintenance, not a feature toggle. The building blocks—ephemeral tokens, CAE, token introspection, JIT admin, and scoped AI connectors—are available. If you still rely on convenience defaults, you’re the low-hanging fruit. Fix the nine settings above, automate discovery and revocation, and stop treating security like a checkbox. We’ll argue about the flashiest vendor next season; the real winners will be the teams who stopped giving attackers long-lived keys to the building.
How often should we audit OAuth apps?
Monthly for mid-market companies; weekly for orgs with >500 seats or regulated data (finance, healthcare). Automate discovery using your IdP and cloud APIs and revoke stale org-scoped tokens immediately.
Is time-bound admin access practical for small teams?
Yes. Use lightweight PAM or your IdP's built-in JIT elevation. It adds an approval step but prevents standing admin accounts attackers love. Make 1–24 hour elevation the default for any sensitive task.
Can AI tools be used safely with sensitive data?
Yes—but only with three controls: (1) connectors scoped and contractually restricted from training on your data, (2) prompt/response logging and retention (90 days+), and (3) strict segmentation so AI assistants never touch legal/security/M&A folders by default.