Sign inJoin the waitlist

ChatGPT for cybersecurity: the useful jobs, the controls, and what OpenAI publishes

Where it earns its place in a security team, the data and certification facts for your vendor review, and the controls that keep it out of the decision path.

Published by AI Knows Us · Updated 12 October 2026

ChatGPT earns its place in a security team on four jobs: explaining unfamiliar code and log output, drafting detection rules and scripts for review, writing policy and awareness material, and speeding up the paperwork around an incident. It is not a control and it does not belong in the decision path of anything automated. OpenAI publishes the commitments that matter for a security review, including that business data is not used to train its models by default, and that the API and business products have been examined under SOC 2 Type 2 with ISO 27001 and ISO 42001 certifications in place. This page covers the good uses, the controls to put around them, and what a cybersecurity vendor should do about being named when buyers ask an assistant.

Where it genuinely helps a security team

  • Reading things fast. Paste an obfuscated script, a stack trace, a PowerShell one liner or a block of firewall logs and ask what it does. This is the highest value use because it compresses the slowest part of triage.
  • Drafting detections. A first pass Sigma, KQL, SPL or YARA rule from a described behaviour, then tuned by an engineer against real data. Treat the output as a draft that an analyst owns.
  • Phishing analysis. Summarising headers, explaining a lure, and listing the indicators worth pivoting on.
  • Policy and documentation. Access control policies, starter playbooks, tabletop scenarios, awareness emails and the plain English version of a standard for a board paper.
  • Incident paperwork. Turning an analyst's notes into a timeline, a customer notification draft and a lessons learned document, with a human signing every word.
  • Secure code review support. Explaining why a pattern is dangerous and suggesting the safer form, alongside your existing scanning tools rather than instead of them.

What OpenAI publishes, for your security review

If you are completing a vendor assessment, these are the statements on OpenAI's own pages rather than hearsay.

TopicWhat OpenAI states
Training on your dataBusiness data from ChatGPT Business, Enterprise, Edu and the API Platform is not used to train its models by default
OwnershipYou own and control your business data, including inputs and outputs where law allows
RetentionRetention is controllable on the Enterprise, Healthcare and Edu products
EncryptionData encrypted at rest with AES-256 and in transit with TLS 1.2 or above
Access controlSAML single sign on, with fine grained control over access and features
Audit and certificationSOC 2 Type 2 examination of controls for security, availability, confidentiality and privacy for the API and ChatGPT business products, plus ISO 27001, 27017, 27018, 27701 and ISO 42001
Regulatory supportSupport for GDPR, CCPA, HIPAA and FERPA compliance, with a Data Processing Addendum and a Business Associate Agreement available

Read the current wording at OpenAI's enterprise privacy page and its security and privacy page, because commitments and certification scope are both updated over time.

The controls to put around it

The risk in a security team is rarely the model. It is the habits that form around it.

  • Use a business tier, not personal accounts. The data commitments above attach to the business products. An analyst pasting client logs into a personal account is operating outside all of them.
  • Write a one page acceptable use note. What may be pasted, what may not, and who to ask. Customer data, credentials, keys, unredacted PII and anything under a client NDA on the no list.
  • Keep it out of automated paths. A model output that triggers a block, a quarantine or a firewall change without a person in the loop turns a wrong answer into an outage.
  • Treat fetched content as untrusted input. Prompt injection is the specific risk when a model reads a web page, an email or a document on your behalf. Instructions hidden in that content are data, not commands, and any workflow that gives a model both untrusted input and a tool that can act needs a human approval step.
  • Verify every artefact before it runs. A generated rule, query or script is reviewed and tested the same way as one written by a junior engineer. Named author, named reviewer.
  • Log the usage. Keep a record of what the team uses it for, so your next audit answers the question with evidence rather than recollection.

Where it does not belong

Three uses are worth ruling out explicitly, because they come up in every team.

  • As the source of truth for a CVE, a patch level or a vendor advisory. Check the vendor's own advisory and the CVE record.
  • As the writer of regulated external communication. Breach notifications carry legal deadlines and defined content, so a lawyer owns the final text.
  • As a substitute for a scanner, an EDR or a log platform. It explains what those tools produce; it does not produce it.

For cybersecurity vendors: being named when buyers ask

Security buyers research by asking questions, and an assistant answering "which endpoint vendor suits a 200 seat firm with a Microsoft stack" is building somebody's shortlist. Four things decide whether you are on it.

  • A specific category claim. "Managed detection and response for mid market financial services in the UK" is an entity a model can file. "Cyber security solutions" is not.
  • Your compliance facts in crawlable text. Certifications with issuing bodies, data residency, the standards you are assessed against. Buyers ask these questions first, and a logo image answers none of them.
  • Independent corroboration. Review platforms, analyst listings, trade press and conference programmes. A claim only on your own domain has one source behind it.
  • A page per buyer question. Title as the question, answer in the first two sentences, one checkable fact in the answer. The structure is in how to write a page an assistant will quote.

Security buyers vet harder than most, and increasingly the first pass happens before anybody contacts you. Somebody asks an assistant which firms handle their situation, and a short list comes back built from whatever the web says about you, including pages you did not write. AI Knows Us runs those questions and shows the sources behind each answer, so a wrong claim about your certifications can be found and corrected at its origin.

This page covers using ChatGPT inside a security team, and visibility for security vendors. For crawler access decisions, see which AI crawlers should you allow.

Questions buyers ask about ChatGPT for cybersecurity

Is it safe to paste logs into ChatGPT?

On a business tier, with customer identifiers and credentials redacted, and under a written acceptable use note, this is a normal working practice in many teams. Redact first, use the business account, and keep client data governed by an NDA out of it.

Does OpenAI train on what my team types?

OpenAI states that business data from its ChatGPT Business, Enterprise, Edu and API products is not used to train its models by default. Personal accounts are governed by consumer settings instead, which is the main reason to standardise the team on a business tier.

Can ChatGPT replace a security analyst?

It changes how an analyst spends an hour rather than removing the role. The judgement calls, the ownership of a decision and the accountability for an action all stay with a named person.

What is prompt injection, in one paragraph?

Content a model reads can contain instructions aimed at the model rather than at you. If that model also holds a tool that can act, such as sending mail or changing a configuration, those instructions become a path into your environment. The control is to treat all fetched content as data and require human approval before any action.

What certifications does OpenAI hold?

Its published list includes a SOC 2 Type 2 examination for the API and ChatGPT business services, ISO/IEC 27001, 27017, 27018 and 27701, and an ISO/IEC 42001 AI management system. Confirm current scope on OpenAI's security page as part of your assessment.

What to do first

  • Move the team onto a business tier and turn on single sign on.
  • Write the one page acceptable use note and name the person who approves exceptions.
  • Pick two tasks to start with, log explanation and detection drafting, and require a named reviewer on both.
  • Review any workflow where a model both reads untrusted content and can act, and add a human approval step.
  • If you sell security, ask five buyer questions in two assistants without naming your company, and save who was named and what was cited.

Read next

AI Visibility Audit Deliverables for a UK Agency: The Five Artefacts, and Why the Zero Column Is One of ThemWhat a British agency should hand a client at the end of an audit, including the findings that did not survive a second reading.A Workflow Standard for AI Search Optimisation: Six Stations, and Who Owns Each OneThe execution sequence an Indian agency can hold a team to, measured against what our own site serves eleven named crawlers.The Multi Engine AI Citation Audit Framework: One Column Per Engine, Never a Blended ScoreA reproducible audit design for an Indian agency, built from 1,260 readings on 16 October 2026 and the 597 that came eleven days before it.A Small UK Agency Against a Large Consultancy on AI Search Optimisation: What the Citation Data Says Each One Can OwnThree British agency domains took dozens of citations each in our readings, and none of them is a large firm. Here is what that result does and does not tell a ten person shop.Build or Buy an AI Visibility Stack: The Five Silent Wrong Answers an In House Build ProducesWhat it actually takes to run a multi engine tracker yourself, measured from the defects in our own build, with rupee prices for the alternative.AI Search Optimisation and SEO in India: What Each One Costs, and What Each Return Is Counted InMost of the production work is already inside your search invoice, two budget lines are genuinely new, and the two results are counted in units that cannot be divided by each other.

Or browse the whole archive.

See what AI says about you.

The first scan is free, it needs no account, and it asks the questions your buyers ask.

Free. No card. We ask 5 real buyer questions on 2 AI apps.

Already sure you want this? Join the waitlist. We open on 31 October 2026. It costs an email address. No card, and nothing is charged. Or talk to a person if you have a question first. Or read how we fix what it finds.