The OSAI exam is a 24-hour proctored red team engagement against an AI-enabled enterprise. You exploit LLM agents, RAG pipelines, multi-agent systems, and the cloud and Kubernetes infrastructure underneath them, then submit a written report that proves every finding. This guide covers the exam format, the full AI-300 syllabus, a week-by-week 90-day study plan, and the mistakes that fail candidates — drawn from public reviews, community feedback, and Macksofy's OSAI bootcamp experience across India and the UAE.
What is the OSAI exam format?
The OSAI exam is a 24-hour proctored practical where you receive VPN access to a realistic AI-enabled enterprise environment. An Offensive Security proctor monitors the session. Your job is to identify and exploit vulnerabilities across the AI application layer and the infrastructure behind it, then write a report that documents every finding with reproducible evidence.
| Dimension | Detail |
|---|---|
| Duration | 24 hours of active testing, followed by a report submission window |
| Delivery | Proctored over VPN — a live proctor monitors your screen and webcam |
| Environment | A multi-host AI enterprise: LLM-powered applications, RAG stores, agent frameworks, and backend infrastructure (cloud IAM, Kubernetes, containers) |
| Tools allowed | Your own Kali/attack VM; community reviews report that candidates are encouraged to use AI assistants during the exam |
| Scoring | Points-based — 75 out of 100 required to pass; partial credit is available for findings with incomplete chains |
| Report | A professional penetration test report with executive summary, methodology, findings and remediation; graded separately from the technical score |
| Credential | OSAI (never expires) + OSAI+ (expires after 3 years, renewable via CPE or a qualifying Offensive Security exam) |
OSAI exam format summary — verify current rules on the Offensive Security AI-300 page before booking
What does the OSAI syllabus cover?
The AI-300 course has 11 learning modules totalling roughly 65 hours. The table below maps each module to its topic areas and an estimated exam weight based on community reports and the depth of lab coverage. These weights are not published by Offensive Security — treat them as directional.
| # | Module | Key topics | Est. weight |
|---|---|---|---|
| 1 | AI Reconnaissance & Threat Modeling | Fingerprinting LLM endpoints, mapping AI architecture from partial intelligence, MITRE ATLAS threat models | 5% |
| 2 | Attacking AI Agents | System-prompt extraction, guardrail evasion, tool-call hijacking, memory poisoning in single-agent systems | 12% |
| 3 | Attacking RAG Pipelines | Retrieval poisoning, indirect prompt injection via documents, access-control bypass in vector stores | 12% |
| 4 | Multi-Agent Systems & A2A Protocols | Agent-to-agent trust exploitation, routing abuse, rogue-agent registration, cross-agent instruction injection | 10% |
| 5 | Attacking Embeddings | Embedding inversion (vec2text), semantic similarity abuse, adversarial embedding manipulation | 8% |
| 6 | Attacking MCP & Tool Surfaces | MCP server vulnerabilities, tool-definition injection, return-value manipulation, schema poisoning | 10% |
| 7 | AI/ML Supply Chain | Pickle deserialization RCE, model file tampering, tokenizer hijacking, Hugging Face module substitution | 8% |
| 8 | AI Infrastructure — Cloud, K8s, Containers | AWS/Azure/GCP IAM from AI footholds, Kubernetes RBAC, container escapes, GPU-node privilege escalation | 12% |
| 9 | Threat Modeling Framework | Scoping AI engagements, risk rating for non-deterministic findings, attack trees for AI-enabled targets | 5% |
| 10 | Capstone Engagement | End-to-end chain: recon through AI app exploitation, infrastructure pivoting, persistence and reporting | 13% |
| 11 | Report Writing | Executive summary, finding documentation for non-deterministic bugs, evidence that reproduces, remediation advice | 5% |
AI-300 module breakdown — estimated exam weights are from community analysis, not Offensive Security
The LearnToExploit Model Attack & Red Teaming course covers the application-layer attack classes behind modules 1–4, 6 and 7, plus five classes OSAI does not teach as separate modules (jailbreaking, output handling, system-prompt extraction). It does not replace modules 5 (embedding inversion), 8 (infrastructure depth) or the capstone. Use it as the first step, not a substitute.
How hard is the OSAI exam compared to OSCP?
OSAI is narrower than OSCP but deeper within its domain. OSCP tests you across web applications, Active Directory, Linux and Windows privilege escalation, and network pivoting — a wide surface with well-documented techniques. OSAI puts you against a smaller target — an AI-enabled enterprise — but the targets are non-deterministic. The same prompt injection that works once may not work the next time you run it, so your methodology has to handle probabilistic output and your report has to include evidence that reproduces.
- 24-hour practical + report
- Broad surface: web, AD, Linux, Windows, pivoting
- Deterministic targets — exploits either work or they don't
- Infrastructure-first: shells, privesc, lateral movement
- Mature ecosystem: 10+ years of community prep material
- 24-hour practical + report
- Narrow surface: AI apps, agents, RAG, infrastructure behind them
- Non-deterministic targets — LLMs give different output each run
- AI-first: prompt injection, agent abuse, then infra pivoting
- New credential: limited community material, fewer mock environments
If you hold OSCP, the infrastructure half of the OSAI exam will feel familiar — cloud IAM, Kubernetes, containers. The AI application layer is the new skill. If you are coming from an AI or ML engineering background, the reverse is true: you understand the models but need to learn pivoting, privilege escalation and professional reporting. The Macksofy OSAI bootcamp is designed for both paths.
What is the OSAI pass rate?
Offensive Security has not published an official OSAI pass rate. Community estimates from early cohorts suggest a first-attempt pass rate around 40–50 percent, comparable to OSCP in its early years. The two most common failure modes reported are: (1) candidates who clear the AI application layer but cannot pivot through the infrastructure, and (2) candidates who find enough vulnerabilities to pass on points but submit a report that the graders reject.
What is the best 90-day OSAI study plan?
Ninety days is the standard access window for the AI-300 Course + Cert bundle. The plan below assumes 15–20 hours of study per week and mirrors the timeline used by Macksofy's OSAI bootcamp. Adjust the pace if you have more or fewer hours available.
| Weeks | Phase | What to do | Hours |
|---|---|---|---|
| 1–2 | Foundations | Complete the LearnToExploit Model Attack & Red Teaming course (12 modules, 9 labs). Get comfortable with prompt injection, stored injection, guardrail evasion, RAG bypass, agent-tool abuse and reporting. | 30–40 |
| 3–4 | AI-300 Modules 1–4 | AI recon & threat modeling, attacking agents, attacking RAG pipelines, multi-agent systems & A2A. Run every lab until you can chain attacks without hints. | 30–40 |
| 5–6 | AI-300 Modules 5–8 | Embeddings, MCP & tool surfaces, AI/ML supply chain, AI infrastructure (cloud/K8s/containers). This is where OSCP-level infrastructure skills matter. | 30–40 |
| 7–8 | AI-300 Modules 9–11 + Challenge Labs | Threat modeling framework, capstone engagement, report writing. If challenge labs are available, complete them. Write a full report for the capstone. | 30–40 |
| 9–10 | Mock Engagement + Report | Run a timed 24-hour mock engagement (Macksofy bootcamp includes one). Write the report under exam conditions. Have a mentor review it. | 30–35 |
| 11–12 | Revision + Exam | Revisit weak modules, re-run labs, refine your reporting template. Book and sit the exam in week 12. | 20–30 |
90-day OSAI study plan — 15–20 hours per week
What tools and resources do you need for OSAI preparation?
The OSAI exam lets you bring your own attack VM. Here is the toolkit that covers the full AI-300 syllabus, from application-layer prompt injection through infrastructure pivoting.
| Category | Tool / Resource | Why you need it |
|---|---|---|
| AI assistant | A reasoning-capable AI subscription (paid tier) | Candidates are encouraged to use AI during the exam; a capable model helps craft injection payloads, analyse non-deterministic output and debug exploit chains |
| AI-300 labs | Offensive Security lab environment (included in the bundle) | The official practice environment; run every lab at least twice |
| LearnToExploit | Model Attack & Red Teaming course (₹10,000) | 9 browser labs mapped to AI-300 modules; covers attack classes the course teaches plus five it does not |
| Python + LangChain | Python 3.11+, LangChain, llama-index, transformers | Scripting custom injection payloads, interacting with agent APIs, automating embedding queries |
| Web proxy | Burp Suite Community or Professional | Intercepting and modifying HTTP traffic to AI endpoints; essential for token manipulation and request replay |
| Container tools | Docker, kubectl, AWS CLI, gcloud CLI | Infrastructure pivoting in modules 8 and 10 — cloud IAM enumeration, Kubernetes RBAC exploitation, container escapes |
| Reporting | A professional template (Offensive Security's or your own) | Start using it from day 1; the exam grades the report as a separate deliverable |
| Note-taking | CherryTree, Obsidian or similar | Structured notes per module; during the exam you need fast recall of techniques across all 11 modules |
OSAI preparation toolkit
Where can you take OSAI training in India, UAE, Singapore, Saudi Arabia and the USA?
OSAI training is available through Offensive Security directly (self-paced via Learn One at $2,749 per year) and through authorised training partners and independent providers. Availability varies by country. For a detailed cost breakdown, see the OSAI cost guide for India.
India
Macksofy runs a mentor-led OSAI bootcamp that pairs the official AI-300 course with the LearnToExploit Model Attack & Red Teaming lab track, a timed 24-hour mock engagement, and mentor report review. Live online across India, classroom batches in Mumbai, and EMI across 3, 6 or 12 months. Other Indian providers include InfoSecTrain (live instructor-led online), Craw Security (classroom batches in Delhi and Noida), and Securium Academy (Zoom live with a free retake claim). Koenig Solutions lists a low-cost prep course (₹22K–28K) but this does not include the Offensive Security exam voucher or labs. See the full comparison of AI security certifications if you are still deciding between OSAI, CAISP and AAISM.
UAE and Dubai
Macksofy runs corporate AI red-team batches for teams in Dubai and Abu Dhabi from its Burj Gate Tower office on Sheikh Zayed Road. Global Knowledge lists the AI-300 e-learning + cert bundle for the UAE and Qatar. For instructor-led delivery, NobleProg offers generic AI red-teaming courses in Dubai (not Offensive Security branded). There is currently no UAE-based Offensive Security authorised training partner for AI-300 — the market is wide open.
Singapore
Ensign InfoSecurity is an official Offensive Security authorised training partner and runs in-person AI-300 cohorts in Singapore with 65+ hours of content, 90 days of lab access, and enterprise cohorts year-round. Ensign is Asia's largest cybersecurity firm. NUS ISS and NTU PACE offer broader AI-and-cybersecurity programmes (SkillsFuture-funded) but neither is OSAI-specific.
Saudi Arabia
Saudi Arabia has almost no OSAI-specific training. NobleProg lists generic AI security courses in Riyadh and Jeddah. Global Knowledge may serve the Saudi market through its Qatar portal. Candidates in KSA will likely train online through Offensive Security's Learn One subscription or through a provider like Macksofy that delivers live online across time zones.
USA
Evolve Security Academy runs an 8-week part-time remote bootcamp ($5,250, includes 90-day voucher and one exam attempt). Applied Technology Academy is an official Offensive Security authorised provider with live online 5-day sessions listed on CISA's NICCS catalogue. Offensive Security's Learn One subscription ($2,749 per year) is the direct option. As of November 2026, OSAI is included in Learn One for the first time.
Is OSAI worth it for your career?
AI red teaming is a new discipline and OSAI is the first hands-on certification from a vendor with a decade of credibility in offensive security. The credential is scarce — launched in March 2026, with a demanding 24-hour practical exam — which means holders stand out in a market where every BFSI, SaaS and GCC team is now shipping copilots and agents but very few have someone who can test them. The economic case depends on your market.
| Market | AI Red Teamer salary range | Demand signal |
|---|---|---|
| India | ₹20–40 LPA | BFSI (RBI AI guidelines), SaaS companies shipping AI features, GCC captives building internal AI platforms |
| UAE | AED 30,000–50,000 / month | DIFC and ADGM fintechs, UAE Cybersecurity Council initiatives, Dubai AI Strategy 2031 |
| Singapore | SGD 8,000–15,000 / month | MAS technology risk guidelines, Smart Nation initiative, regional SOC expansion |
| Saudi Arabia | SAR 25,000–45,000 / month | Vision 2030 digital transformation, NDMO data governance requirements, NEOM and giga-project security teams |
| USA | $150,000–250,000 / year | SEC cybersecurity disclosure rules, AI executive order, enterprise AI adoption across Fortune 500 |
AI red teamer salary ranges by market — from public job boards, November 2026
The AI pentesting and red teaming service lines at Macksofy are built around this skill set. If you are deciding whether OSAI is the right next credential after OSCP or OSEP, the certification comparison guide covers the full landscape, and the OSCP-to-OSAI career path maps the progression.
What mistakes do OSAI candidates make?
These are the failure patterns reported by candidates and confirmed by Macksofy's mock-engagement reviews. Avoid them and your first-attempt odds improve significantly.
- **Treating LLMs as deterministic targets.** The same injection prompt may work three times and fail on the fourth. Candidates who run a payload once, see it succeed and move on often find that their evidence does not reproduce during grading. Run every attack multiple times and document the success rate.
- **Skipping report writing until exam day.** The exam grades the report separately. Candidates who spend 23 hours exploiting and 1 hour writing fail on the report even when their technical score is above 75. Start your reporting template on day 1 of preparation.
- **Not using AI tools.** This is the one Offensive Security exam where AI assistants are encouraged. Candidates who try to craft every injection payload by hand waste hours. Use a reasoning model to generate, iterate and debug payloads.
- **Ignoring the infrastructure layer.** Modules 8 and 10 require OSCP-level skills: cloud IAM enumeration, Kubernetes RBAC exploitation, container escapes. Candidates from AI or ML engineering backgrounds often stop at the application layer and miss the points behind it.
- **Skipping LearnToExploit or equivalent foundational practice.** The AI-300 course assumes you can already inject a prompt and understand how agents process tool calls. Candidates who jump directly into AI-300 without hands-on LLM attack experience spend their first weeks learning basics instead of building depth.
- **Not practising under time pressure.** Twenty-four hours sounds generous until you are debugging a non-deterministic chain at 3 AM. Run at least one timed mock engagement before booking the exam — the Macksofy bootcamp includes one.
- **Poor scoping and prioritisation.** The exam environment has more targets than you can fully exploit in 24 hours. Candidates who try to chain everything end up with many partial findings and few complete ones. Identify the high-point targets early, chain them completely, then move to the next.
Mentor-led AI-300 preparation with the LearnToExploit lab track, a timed 24-hour mock engagement, and report review. Live online across India, classroom in Mumbai, corporate batches in Dubai and Abu Dhabi. EMI available.
