// guides

How to Choose a Penetration Testing Company: 12 Vendor Questions

How to choose a penetration testing company: security lead vetting pentest vendors against twelve questions

To choose a penetration testing company, check three things you can verify before you sign: who will actually do the testing, whether every finding ships with a reproducible proof of exploitation, and whether the methodology covers your real surface (cloud control plane, APIs, AI agents). Brand and price come after those, because those three separate a real test from a resold scan.

The short answer

  • Judge vendors on named testers, a redacted sample report with a retest report, and a methodology document mapped to PTES, OWASP WSTG/ASVS and MITRE ATLAS for AI.
  • Require a written rule that every reported finding includes a working, reproducible proof. Scanner plugin IDs in a 'manual' report disqualify the vendor.
  • Agree on retest cycles, critical-finding escalation SLA, insurance, liability cap and data deletion before kickoff, not after the report arrives.
  • If a vendor cannot name testers, will not show a report, or answers AI and cloud questions with 'we run a tool and review it', end the call.

Key takeaways

  • Verify named testers, a redacted sample report with a retest report, and a written methodology before signing.
  • Hands-on certifications (OSCP, OSWE, CREST CRT, GIAC GPEN) are a minimum bar. CISSP shows governance knowledge, not exploitation skill.
  • Cloud pentests must demonstrate IAM attack paths. AI tests must map to the OWASP LLM Top 10 (2025) and MITRE ATLAS and report an attack success rate.
  • PTaaS is a delivery model, not a different test: apply the same tester, proof and methodology checks.
  • Put retest cycles, escalation SLA, insurance, liability cap and GDPR Article 28 data terms in the contract before kickoff.
  • DORA TLPT is an intelligence-led red team under TIBER-EU, not a web app pentest.

How do I choose a penetration testing company?

Pick the vendor that can show you three things in writing: the named people who will test you, a sample of their evidence standard, and a methodology that covers the surface you care about. Logo, price and platform UI come second.

A pentest buys skilled human judgment over a fixed number of days. The deliverable is only as good as the person holding Burp Suite and whatever they decided to chase on day three. Part of the market sells something cheaper: Nessus or Qualys output, lightly edited, in a template with “manual penetration test” on the cover. Both products produce a PDF.

So read the findings. A resold scan reads like configuration hygiene. A manual test reads like an attack. These are the fingerprints of a resold scan:

  • Scanner plugin IDs or verbatim plugin descriptions in a report labeled manual.
  • No chained findings. Every issue stands alone, and nothing leads anywhere.
  • Almost every finding is a configuration item: TLS ciphers, missing headers, version banners.
  • No attack narrative showing how the tester moved from foothold to impact.
  • No authenticated testing across roles or tenants, so no BOLA, no IDOR, no business-logic flaws.
  • Boilerplate remediation that would read the same for any client.

Before evaluating vendors, decide which product you are buying:

Engagement What it proves Typical effort Fits when
Automated scan Known CVEs and misconfigurations exist Hours Continuous hygiene, asset inventory
Manual pentest Specific weaknesses are exploitable, with proof 5-20+ tester-days Release gates, compliance, new features
PTaaS subscription Manual findings against a changing target, delivered through a platform Recurring tester-days per cycle Frequent releases, need for continuous evidence
Red team An objective is reachable, and whether detection catches it 4-8 weeks Mature programs testing detection and response
Bug bounty Researchers found something they could get paid for Ongoing Broad public surface, triage capacity in place

PTaaS is a delivery model, not a different kind of test. Hold it to the same three checks, because some subscriptions are a scanner dashboard with occasional human review. A pentest priced as a pentest but built from scanner output is the most expensive scan you will ever buy.

The twelve questions below expose the difference in one scoping call. Each comes with the answer that should end it.

Who will hold the keyboard, and can they prove their work?

Question 1: Who exactly will test us? Ask for names in the SOW, whether each person is an employee or a contractor, and whether any work is subcontracted, including offshore. Add a clause requiring your written consent before any substitution. Ask about each tester’s public output (CVEs, talks, tools, write-ups), and get two references from clients of similar size and regulatory scope.

Treat certifications as a minimum bar. OSCP, OSEP, OSWE, CREST CRT and GIAC GPEN or GWAPT all require hands-on exploitation to pass. CISSP and CISA are governance credentials: useful for whoever writes your policy, silent on whether someone can exploit a deserialization bug. Answer that ends the call: “We assign from our pool after signature,” with no right to approve the person.

Question 2: Can we see a redacted sample report from a comparable scope, plus a retest report? Every credible firm keeps a sanitized sample ready. A good report contains:

  • An executive summary written in business impact, not CVE counts.
  • An attack narrative showing how findings chained together.
  • Reproduction steps and raw request/response evidence for each finding.
  • A severity vector plus a business-context rating.
  • Remediation written for your stack, not generic OWASP links.
  • Low and informational findings. Senior testers show judgment in what they decline to escalate. A sample with only Criticals is marketing.

The retest report matters as much as the main one, because auditors and enterprise customers accept it as evidence that issues were closed. Answer that ends the call: “Our reports are all under NDA.”

Question 3: What is your methodology, and what standard does it map to? Ask for the document itself. PTES defines seven phases, from pre-engagement through exploitation and post-exploitation to reporting. A vendor that cannot describe its pre-engagement process (scope, rules of engagement, escalation) or what it does with access after exploitation is skipping half the work.

NIST SP 800-115 supplies planning and reporting discipline. OWASP WSTG supplies test cases: v4.2 organizes roughly 100 of them into 12 categories, and a v5 restructure has been in progress, so ask which version the plan follows. OWASP ASVS defines what passing means; ask which level (L1, L2 or L3) of ASVS 5.0 you will be measured against.

For APIs, expect the OWASP API Security Top 10 (2023 edition) with BOLA and BFLA hunted across every object ID. For mobile, expect MASVS. Grey-box testing should be the default, meaning credentials, an architecture walkthrough, and test accounts across roles and tenants. Without multiple roles, authorization testing is guesswork. Answer that ends the call: “Our methodology is proprietary,” and no document follows.

Is a cloud configuration review the same as a cloud pentest?

No. A configuration review lists what is misconfigured. A cloud pentest proves which identity can reach which crown jewel, and by what path. Many quotes blur the two.

Question 4: How do you test the cloud control plane? A real answer covers IAM attack-path analysis (privilege escalation chains, trust relationships, role chaining, cross-account pivots) plus the data-plane and application testing on top. The vendor should name tools without hesitation: Pacu and CloudFox for AWS offense, ScoutSuite and Prowler for posture, BloodHound/AzureHound, ROADtools and AADInternals for identity paths. Then they should explain the manual work that follows the tool output.

Ask which AWS escalation patterns they routinely demonstrate. Good answers include iam:PassRole combined with lambda:CreateFunction, policy version rollback, STS role chaining and abuse of cross-account trust. Those are attack-path findings. “S3 bucket is public” is a checklist item. Our AWS penetration testing of IAM attack paths is scoped around these chains for that reason.

Answer that ends the call: “We run Prowler and review the results.” That is a CSPM report sold at pentest rates.

Cloud pentest vs configuration review: IAM attack path through cloud accounts to a protected vault
A config review lights up warnings. A pentest traces the path to the vault.

How do I evaluate a vendor's LLM and AI-agent testing?

Ask for an AI test plan mapped to the OWASP Top 10 for LLM Applications (2025) and MITRE ATLAS, a defined attack corpus, and a measured attack success rate. “We tried some jailbreaks” is not a test plan.

Question 5: How do you test our LLM, MCP and RAG components? Check that the vendor works from the current list. The 2025 edition includes LLM06 Excessive Agency, LLM07 System Prompt Leakage and LLM08 Vector and Embedding Weaknesses; a deck quoting the 2023 list word for word is built on old material. Ask whether their threat model follows NIST AI 100-2, which separates integrity, availability, privacy and abuse attacks. Jailbreaks cover only one of those.

For Model Context Protocol servers, the plan should cover tool poisoning through malicious tool descriptions, “rug pull” redefinition of a tool after approval, confused-deputy paths where the server acts with the user’s authority, token passthrough and session hijacking. The MCP project publishes security guidance alongside the specification; ask whether they test against it.

For RAG, expect indirect prompt injection through poisoned retrieved documents, cross-tenant retrieval leakage, embedding inversion and system-prompt leakage. Agent testing needs multi-turn, stateful chains that end in a tool call, not single prompts. Garak, PyRIT and promptfoo are reasonable starting tools. Ask whether the corpus is reused unchanged across clients, because a corpus your developers have already tuned against inflates your pass rate.

The EU AI Act is also pushing AI red-team clauses into contracts: Article 15 sets cybersecurity requirements for high-risk systems, and Article 55 requires adversarial testing of general-purpose AI models with systemic risk. Our AI agent penetration testing is scoped this way, and the AI agent penetration test scope checklist lists the inputs to give any vendor. Answer that ends the call: “Our scanner has an AI module.”

LLM AI agent penetration testing vendor evaluation: tampered tool chain and poisoned documents feeding an agent
AI testing means tool chains and retrieved documents, not single jailbreak prompts.

How should findings be proven, rated and escalated?

Every reported finding should include a working, reproducible proof, and the contract should say so. Unvalidated findings cost your engineers hours and wear down trust in the report until nobody reads it.

Question 6: What is your proof-of-exploitation standard and false-positive policy? Look for request/response pairs, timestamped screenshots and steps a developer can replay. Theoretical findings should be labeled theoretical, not padded into the severity count.

Severity needs more than CVSS. CVSS v4.0 (November 2023) produces a technical score with no business context. Mature vendors add a business-impact rating and, for known CVEs, EPSS exploit probability or a CISA SSVC decision. A medium SSRF that reaches cloud metadata on an internet-facing service should be fixed before a critical on an isolated lab host. Answer that ends the call: “The scanner rated it critical.”

Question 7: What is your critical-finding escalation SLA, and what are your stop conditions? A tester who finds a critical on day two should call you within hours. A critical finding that sits unreported is a live exposure you paid to find out about. Get in writing who gets the call, within how many hours, and what happens next.

Ask what the tester does with domain admin or cloud admin once they have it. Post-exploitation should go only as far as the rules of engagement allow, then stop. Name an emergency-stop contact on both sides. Answer that ends the call: “Everything goes in the final report.”

What should the quote and contract cover?

Size the work in tester-days per asset and compare quotes on that basis, not the headline price. Typical ranges: 5-10 days for a mid-complexity authenticated web app, 10-20+ days for a complex API estate or cloud account, 4-8 weeks for a full red team.

Question 8: What does a “day” mean in your quote? Ask whether it means eight hours of one named tester, how much of the time is reporting, and whether scoping calls count against it. A 10-day quote with 4 days of reporting is a 6-day test.

Scope Indicative US/EU range
Single web application $8k-$30k
Cloud attack-path review $10k-$40k
AI/LLM red-team engagement $25k-$75k
Full red team $50k-$150k+

These are rough ranges. Blended day rates commonly fall between about $1,200 and $2,500. A quote far below the range has usually removed the manual hours, and those hours are what you are paying for.

Question 9: How many retest cycles, within what window, and what counts as fixed? One retest of confirmed findings should be included free or at nominal cost, usually valid for 30-90 days and limited to the original findings. “Fixed” should mean the original proof no longer works and no obvious bypass does either. A vendor who bills a retest as a new engagement is telling you how they intend to treat you after delivery.

Question 10: What legal and insurance cover do you bring? Require a signed authorization letter naming your legal entity, the testers, the testing window, source IPs, prohibited techniques (DoS, destructive social engineering) and the emergency contact. Ask for certificates of professional indemnity/E&O and cyber liability insurance. Limits of $1M-$5M are a common enterprise reference point; negotiate for your own risk.

Read the liability cap closely, since it is often set at 12 months of fees. If the tester takes production down, that clause decides who pays.

Question 11: How do you handle our data? Pentest evidence regularly contains personal data and live credentials, which makes the vendor a processor under GDPR Article 28. You need a DPA, a defined retention period, a certificate of deletion, a sub-processor list and disclosure of any transfers outside the EU. Answer that ends the call: “We keep everything indefinitely for quality purposes.”

Does a pentest satisfy PCI DSS, SOC 2, ISO 27001 or DORA?

A properly scoped pentest can satisfy all four, but each framework expects different evidence, and DORA expects a different exercise altogether.

  • PCI DSS v4.0.1 Requirement 11.3: internal and external testing at least annually and after significant changes, plus segmentation testing every six months for service providers. The future-dated v4.0 requirements became mandatory on 31 March 2025.
  • SOC 2: does not require a pentest by name, but auditors commonly request one as evidence under CC4.1 and CC7.1, and so do customer questionnaires.
  • ISO/IEC 27001:2022: pentest evidence usually maps to Annex A 8.8, management of technical vulnerabilities.
  • DORA: applies from 17 January 2025 and requires Threat-Led Penetration Testing at least every three years for designated financial entities, following TIBER-EU. TLPT is an intelligence-led red team, not a web app test, and most vendors cannot deliver it.

Question 12: What will you hand our auditor? Ask for an attestation letter template and confirm it names the scope, dates and retest status your framework requires.

Red flags and the final scorecard

Any one of these is enough to walk away:

  • No named testers, or a “certified team” made up mostly of CISSPs.
  • No sample report, or a sample containing only Criticals.
  • Scope taken from the vendor’s own asset discovery instead of your architecture and threat model.
  • Black-box testing offered as the default for an authenticated application.
  • An “AI testing” line item with no corpus, no attack success rate and no OWASP LLM or ATLAS mapping.
  • A cloud pentest that is really a CSPM export.
  • Retests billed as new engagements, or no escalation SLA.
  • No DPA, no insurance certificate, or an uncapped indemnity in the vendor’s favor.

To make the decision defensible, score each finalist: testers and their evidence 25%, sample and retest reports 20%, methodology and coverage of your surface 20%, validation and severity practice 15%, commercial and legal terms 20%. Compare price only between vendors who clear that bar.

How to choose a penetration testing company: resold scan report versus manual pentest attack chain
Identical pages versus a chain of evidence: the difference shows in the findings.

FAQ

How do I tell if a pentest is just an automated scan?

Look for scanner plugin IDs, findings that are all configuration items, no chained exploits and no evidence of authenticated testing across roles. A manual pentest report includes an attack narrative and reproducible proof for each finding.

Is OSCP or CISSP better for a penetration tester?

For hands-on testing, OSCP. It is a practical exploitation exam. CISSP is a governance certification and does not show that the holder can exploit anything.

How much does a penetration test cost in the US and EU?

As rough ranges, a single web app runs about $8k-$30k, a cloud attack-path review $10k-$40k, an AI/LLM red team $25k-$75k and a full red team $50k-$150k or more. Compare quotes by tester-days per asset rather than headline price.

Should a retest be included in a pentest quote?

Yes. One retest of the original findings, valid for 30-90 days after the report, is standard. A vendor who charges a retest as a new engagement is a commercial red flag.

Is a cloud configuration review the same as a cloud penetration test?

No. A configuration review lists misconfigurations against a benchmark. A cloud penetration test proves which identities can escalate privileges or pivot across accounts to reach sensitive data, and shows the path.

Should I buy a point-in-time pentest, PTaaS or a bug bounty?

Buy a point-in-time pentest for release gates and compliance evidence, PTaaS when you ship often and need recurring manual coverage, and a bug bounty only for a broad public surface with triage capacity in place. Hold every option to the same tester and proof standards.

If you want to run these twelve questions against us, we will name the testers before you sign, share a redacted report and retest report, and show how we scope API penetration testing for BOLA and BFLA, cloud and AI-agent work.



// 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