// guides

Assumed Breach Testing: The Best-Value Security Engagement

Assumed breach testing concept: a compromised identity node spreading access across a corporate permission graph

Assumed breach testing skips initial access and starts from a compromised laptop, cloud identity or agent token, then measures how far that foothold reaches and how fast your team contains it. For most companies, that is the question that decides breach impact, which makes it the engagement that moves the most risk per dollar.

The short answer

  • Assumed breach testing starts from a realistic compromised position (laptop, cloud identity or agent token) instead of earning initial access.
  • It measures what decides loss: reachability, blast radius, detection coverage and time to contain.
  • Scope it around one crown-jewel business scenario with a credible initial access narrative and least-privilege start.
  • For most companies it delivers more risk-changing findings per dollar than a full red team, but it cannot test perimeter or phishing resistance.

Key takeaways

  • Assumed breach testing starts from an agreed compromised position and measures reachability, blast radius, detection and containment.
  • Adversary-in-the-middle (AiTM) session theft, access brokers and leaked non-human credentials make initial access the cheap part of an attack.
  • Scope backward from one crown-jewel scenario, with a documented, intelligence-backed initial access narrative and the least privilege an attacker would plausibly hold.
  • Judge the engagement on measured detection and eviction times and hop counts to crown jewels, not on reaching admin.
  • AI agent tokens need their own scope: tool inventory, injectable data sources, memory persistence and missing human approval.

What is assumed breach testing and how does it work?

Assumed breach testing is an adversary-style engagement in which the testers receive a realistic compromised starting position instead of earning one. The position is agreed in writing: a standard corporate laptop with EDR running, a low-privilege cloud identity with MFA already satisfied, or a leaked API or agent token.

From there, the work is pure post-compromise testing. The team enumerates what the identity can see, escalates where it can, moves laterally into SaaS, cloud and CI/CD, sets up persistence and tries to reach agreed crown-jewel data. The defenders either respond live or are measured afterward against the telemetry the attack produced.

The name comes from Zero Trust. NIST SP 800-207 (August 2020) treats assume breach as a core principle, usually listed alongside “never trust, always verify” and least privilege: treat the environment as hostile, minimize blast radius, segment internally and use analytics to detect. CISA’s Zero Trust Maturity Model v2.0 (2023) carries the same premise across identity, devices, networks, workloads and data.

Assume breach is the design principle. Assumed breach testing checks whether the design actually holds.

How is assumed breach different from a red team or a pentest?

Assumed breach drops the initial-access phase that a full red team includes and adds the impact, detection and response measurement that a targeted pentest leaves out. They answer different questions and are not substitutes.

Engagement What it proves Typical effort Buy it when
Full-scope red team Initial access resistance plus post-compromise impact and detection 6-10+ weeks Mature controls need an end-to-end, stealthy adversary test
Assumed breach Reachability, blast radius, detection and containment from a defined foothold 2-4 weeks for a mid-size scope You need to know what one compromised identity costs you
Targeted pentest Vulnerability classes in a scoped app, API or network 1-3 weeks A system is new, changed or in compliance scope
Breach and attack simulation Whether known techniques trigger existing controls, continuously Ongoing, automated You want regression checks between human tests
Tabletop Decision-making, escalation and communications Hours to days Leadership and legal workflows are the gap

Success criteria differ too, depending on where the estate lives. On-prem, a pentest traditionally ended at domain admin. In the cloud there is no domain admin, only a permission graph. The useful question becomes: from this identity, which roles, buckets, databases and SaaS tenants are reachable, and in how many hops? BloodHound and AzureHound exist because that graph is where the real attack path lives.

The limit is just as clear. Assumed breach cannot tell you whether your perimeter, phishing resilience or external attack surface would stop someone getting in. If that is your question, pair it with external testing and phishing simulation, or buy full-scope red team operations when a mature program needs the whole chain exercised unannounced.

Assumed breach testing maps hops from one compromised identity to crown-jewel data in a cloud permission graph
In cloud there is no domain admin, only a graph of reachable roles and data.

Why is assumed breach the best value per dollar?

Initial access is now cheap for attackers and expensive for testers to prove, while post-compromise impact is where losses actually come from. Assumed breach spends the budget on the second part.

Three things made initial access a commodity:

  • Adversary-in-the-middle (AiTM) phishing. Reverse-proxy kits in the evilginx class, and commercial kits such as Tycoon2FA, capture session cookies and OAuth tokens after the victim completes MFA. The attacker inherits an authenticated session. A stolen session is, functionally, a valid identity: exactly the state assumed breach starts from.
  • Initial access brokers. Footholds sell on criminal markets, priced by privilege and target value. Ransomware crews buy access rather than build it.
  • Secrets sprawl and non-human identities. Service principals, CI/CD roles, bots and agents outnumber human accounts in most enterprises. GitGuardian’s 2025 State of Secrets Sprawl report counted 23.8 million new secrets leaked to public GitHub in 2024. Each leaked key is a pre-built foothold.

Now the clock. CrowdStrike’s 2025 Global Threat Report put average eCrime breakout time at 48 minutes, with the fastest observed case at 51 seconds. If your detection and response cycle is slower than that, the attacker spends the second hour in your crown jewels.

A full red team spends a large share of its budget on reconnaissance and access you already know can be bought. Assumed breach spends that money on the part that decides whether an incident becomes a ticket or a disclosure.

How do you scope an assumed breach engagement?

Scope it backward from one business-impact scenario, then fix a realistic starting position in writing. Scope by asset inventory and the test drifts into a rebadged vulnerability scan.

Pick a credible starting position

Use the least-privileged position a real adversary would plausibly hold under your threat model, not a convenient one. Starting as Global Admin or domain admin tests nothing.

Starting position Realistic configuration What it reveals
Compromised endpoint Patched Windows or macOS laptop, standard user, healthy EDR, Entra ID joined, live browser SSO to email and SaaS Credential exposure, local escalation, EDR efficacy, lateral movement, browser token theft
Compromised cloud identity Low-privilege Entra ID or Okta user with MFA satisfied or a session cookie, or a limited AWS IAM access key Directory recon, consent and role abuse, permission-graph reachability, cloud audit coverage
Non-human identity Service principal, workload identity or CI/CD role with its real permissions Standing privilege, secret reuse, revocation gaps, unmonitored automation
Compromised agent token API or OAuth token for an LLM agent or MCP server, for example read on a vector store plus write to a ticketing or CRM tool Tool-chain escalation, injection paths, exfiltration through tool side effects

Choose the one your threat intelligence or incident history says is most likely. A finance firm hit by AiTM phishing last year should start from a stolen SSO session. A software company with a busy CI/CD pipeline should start from a leaked pipeline credential.

Settle the rules before signing

  • The scenario. Name the crown-jewel process: payment release, claims adjudication, source code integrity, customer PII.
  • Initial access narrative. A documented, intelligence-backed story of how the attacker got there. It defeats the inevitable “that would never happen here” objection when findings reach the board.
  • Blue team model. Fully unannounced, a defined response window, or purple team with live detection engineering.
  • Out of bounds. Excluded systems, techniques and data, plus an emergency stop contact.
  • Cloud provider terms. AWS, Azure and Google Cloud each publish testing policies with restricted categories. Check the current terms and record them in the scope document.

Run it in production, with guardrails

It belongs in production, because that is where your real permissions and telemetry live. Three conditions make that safe:

  • Authorization. Written approval naming systems, accounts, time windows and source infrastructure. That is your cover under the US Computer Fraud and Abuse Act, the UK Computer Misuse Act 1990 and EU national laws implementing Directive 2013/40/EU.
  • Data. Never exfiltrate real customer PII to prove a point; doing so can itself be a notifiable incident under GDPR. Use canary or synthetic records.
  • Cleanup. Every app registration, consent grant, mailbox rule and planted memory entry gets logged and removed, and the client verifies removal.

What should an assumed breach test prove?

It should prove four things with evidence: how far the foothold reaches, what it can touch, whether you saw it, and how fast you stopped it. “We got admin” is not one of them.

The objectives that actually test impact:

  • Privilege escalation across identity tiers (Tier 2 to Tier 1 to Tier 0)
  • Tenant-wide reconnaissance with graph tooling such as BloodHound, AzureHound and ROADtools
  • Lateral movement into SaaS, cloud consoles and CI/CD
  • Persistence: new credentials, app registrations, OAuth consent grants, mailbox rules, agent memory
  • Reaching crown-jewel stores, then a controlled exfiltration of canary records

Map every step to MITRE ATT&CK so defenders share the vocabulary: T1078.004 Cloud Accounts, T1528 Steal Application Access Token, T1098.001 Additional Cloud Credentials, T1558.003 Kerberoasting, T1530 Data from Cloud Storage, T1567 Exfiltration Over Web Service.

The metrics that justify the spend:

  • Time to detect, acknowledge and contain, per technique
  • Share of in-scope ATT&CK sub-techniques with a validated detection
  • Identities and hosts reachable from the start position
  • Hops from start to each crown jewel
  • Records and systems inside the blast radius
  • Time to eviction, including token revocation

If the logs needed to measure detection were never retained (Entra ID sign-in and audit logs, Okta system log, CloudTrail, SaaS audit, LLM gateway and tool-invocation logs), that is a finding in itself.

The highest-yield setup is the purple-team variant. Testers and detection engineers work side by side, write Sigma rules or custom detections against the actual technique, verify them live, then replay the chain in a walkthrough.

Recurring findings also make a sanity check on any quote: over-privileged service accounts, standing admin, shared local admin passwords, unmanaged devices with valid sessions, missing egress control, secrets in pipelines, excessive OAuth scopes and no token revocation path. A sample report showing missing patches but no blast-radius or detection data is a discount pentest.

Assumed breach metrics: measuring detection and containment time as defenders track a post-compromise attacker
The deliverable is measured detection and eviction time, not a trophy screenshot.

How do you run an assumed breach test against AI agents and MCP tokens?

Start from the agent’s token and ask what an attacker holding it, or steering it, can make the agent do. Treating it like a plain API key misses the point: the agent’s real privilege is the sum of its tools, its data sources and the absence of human approval.

The Model Context Protocol, published by Anthropic in November 2024, is now the default connector between agents and tools. Its documented risk themes map directly onto post-compromise objectives: tool poisoning through descriptions, confused-deputy patterns, token passthrough, overly broad scopes and implicit trust in locally run servers.

The Agent2Agent (A2A) protocol, which Google announced in April 2025 and later moved to the Linux Foundation, adds a lateral path between agents. The test question: can agent A talk trusted agent B into acting on its behalf?

Agent-specific cases that belong in scope:

  • Token scope enumeration and tool inventory discovery from the token’s perspective
  • Indirect prompt injection through a trusted source, including RAG poisoning
  • Cross-user or cross-tenant retrieval from the vector store
  • Persistence through agent memory
  • Exfiltration through tool side effects: email, tickets, webhooks, commits
  • Escalation by chaining low-risk tools into a high-risk action
  • Destructive actions with no human in the loop

Anchor findings to the OWASP Top 10 for LLM Applications (2025), especially LLM01 Prompt Injection, LLM06 Excessive Agency and LLM08 Vector and Embedding Weaknesses. For the tool-chain layer, MCP server and tool-chain security testing covers schema tampering and credential scope, and our guide on how to test excessive agency in AI agents walks through the escalation cases.

Does assumed breach testing satisfy DORA TLPT, PCI DSS or NYDFS?

It can support all three, but it satisfies none of them automatically. Compliance frameworks dictate scope and process; a risk-driven test follows your most dangerous business path. Know which one you are buying.

  • DORA TLPT. DORA applies from 17 January 2025 and requires threat-led penetration testing at least every three years for designated financial entities, built on TIBER-EU, against critical or important functions in live production. TIBER-EU allows control-team “leg-ups” when a phase stalls, which resembles an assumed start, but TLPT still expects an intelligence-led end-to-end test. Assumed breach between cycles is preparation, not a replacement.
  • CBEST. The Bank of England’s intelligence-led framework follows the same logic for UK financial firms.
  • PCI DSS v4.0. Requirement 11.4 mandates internal and external penetration tests at least every 12 months and after significant changes, plus segmentation testing. An internal assumed breach test scoped to the cardholder data environment can evidence the internal test.
  • NYDFS Part 500. Covered entities need a penetration testing program, and assumed breach results feed it directly.
  • Notification clocks. Blast-radius evidence prepared in advance speeds two decisions: SEC Form 8-K Item 1.05, due four business days after a company determines an incident is material, and GDPR Article 33’s 72-hour notification to the supervisory authority.

Whatever drives the purchase, a one-off loses value fast. Identity graphs, OAuth grants and agent fleets change weekly. Run it at least annually against a rotating crown-jewel scenario, and use automated control validation in between to catch detection regressions.

Three assumed breach starting positions: compromised endpoint, cloud identity and AI agent token
Pick the foothold your threat model says is most likely, at the least privilege.

FAQ

What is assumed breach testing?

Testers get a realistic compromised start, such as a standard laptop, a cloud identity with MFA satisfied or a leaked agent token, and measure reachability, blast radius, detection and containment. It skips the work of earning initial access.

Is assumed breach cheaper than a red team?

Usually. A mid-size assumed breach engagement typically runs 2-4 weeks of testing effort versus 6-10 or more for a full red team, because reconnaissance and initial access are removed. Pricing varies widely with scope and region.

What is the difference between assume breach and assumed breach testing?

Assume breach is a Zero Trust design principle in NIST SP 800-207: build as if the attacker is already inside. Assumed breach testing is the offensive engagement that checks whether that design actually limits blast radius and enables detection.

Should the blue team know about an assumed breach test?

Not necessarily the details, but the engagement needs a defined response window or a purple-team phase plus a replay walkthrough. A red-only run produces findings with no evidence about detection or containment.

Can assumed breach testing replace DORA TLPT?

No. DORA TLPT requires an intelligence-led, end-to-end test built on TIBER-EU with supervisory involvement. An assumed breach test is strong preparation between TLPT cycles and can cover business paths the regulator's scope leaves out.

How often should you run an assumed breach test?

At least annually with a rotating crown-jewel scenario, and after major identity, cloud or agent deployment changes. Automated control validation between tests catches detection regressions.

agent::offense runs assumed breach assessment engagements from endpoint, cloud identity and agent-token starting positions, with measured detection times and validated fixes in the report.



// get started

Work with AgentOffense

Tell us about your target and goals. We’ll reply with scope and a fixed-price quote — usually within one business day.

./request_engagement