[ AI Security ]
Close the Gaps Attackers Find in AI Apps
We test and harden LLM applications and agents against prompt injection, data leakage, tool abuse and supply chain risks. Every finding comes with a reproduction, a fix and a retest.
- Certified information security
- ISO 27001
- Certified information security
- Top 10 for LLM apps as the test baseline
- OWASP
- Top 10 for LLM apps as the test baseline
- Signed before any system access
- NDA
- Signed before any system access
- Weeks for a review or an agent assessment
- 1 to 4
- Weeks for a review or an agent assessment
Layer 1
Edge
Auth and SSO / Rate limiting / WAF
Layer 2
AI gateway
Injection classifier / PII redaction / Output filters
Layer 3
Application and agent
Prompt isolation / Tool allow lists / Confirmation steps
Defense in Depth for LLM Applications
Each layer assumes the one above it can fail. An attacker has to get past all of them, not just one prompt.
Layer 1
Edge
Controls applied before a request reaches the model.
- Auth and SSO
- Rate limiting
- WAF
- Bot detection
Layer 2
AI gateway
Inspects every prompt and response in both directions.
- Injection classifier
- PII redaction
- Output filters
- Token budgets
Layer 3
Application and agent
Keeps instructions, context and tools separated.
- Prompt isolation
- Tool allow lists
- Confirmation steps
- Sandboxes
Layer 4
Data and secrets
Limits what the AI can read and with which identity.
- Permission-aware retrieval
- Secrets vault
- Short-lived tokens
- Encryption
Layer 5
Security monitoring
Detects attacks and records evidence.
- SIEM integration
- Attack alerts
- Audit log
- Incident playbook
AI Security Work We Do Most Often
AI features add attack paths that classic penetration tests miss. These are the engagements teams ask for most.
Pre-launch LLM penetration test
A structured test of your chatbot, copilot or agent before release, covering injection, jailbreaks, data exposure and abuse of connected tools.
Launch with known risks fixed and documented
RAG data leakage review
Check whether retrieval can return documents a user should not see, and fix permission filtering at the index and query level.
Users only retrieve what they are allowed to read
Agent and MCP tool hardening
Review every tool an agent can call, tighten scopes, add confirmation for destructive actions and block chained calls that escalate privileges.
A compromised prompt cannot become a compromised system
Model supply chain review
Audit open-weight models, datasets, packages and MCP servers you depend on for tampering, unsafe serialization and license issues.
Third-party AI components are tracked and verified
Guardrail implementation
Input and output filters, PII detection, topic limits and rate limiting added to an existing AI app without rewriting it.
Unsafe requests and responses are stopped at the edge
Security questionnaire support
Technical evidence and clear answers for the AI sections of enterprise security reviews, backed by real test results.
Security reviews close with evidence, not promises
How We Secure an AI Application
We treat the model as an untrusted component and design controls around it, the same way you would treat user input.
- Prompt injection defenseSeparation of instructions from untrusted content, spotlighting, classifier checks and least-privilege tools, layered because no single filter is enough.
- Data leakage preventionDocument-level access control in retrieval, PII redaction, output scanning and limits on what system prompts and context can reveal.
- Secrets handlingNo keys in prompts or client code. Credentials live in a vault, are injected server-side and rotated, and agents get short-lived tokens.
- Access controlUser identity passed through to every tool and data source, so the AI can never act with more rights than the person using it.
- Sandboxed executionCode interpreters and browser agents run in isolated containers with network egress rules and resource limits.
- Adversarial testingAutomated attack libraries plus manual testing by engineers, mapped to OWASP LLM categories and MITRE ATLAS techniques.
- Detection and responseLogging of prompts, tool calls and blocked attempts, with alerts on attack patterns and a playbook for AI incidents.
- Abuse and cost protectionRate limits, token budgets and bot detection that stop denial of wallet attacks and scraping of your AI endpoints.
Every Finding Mapped to a Framework Your Team Knows
Tested against
- OWASP Top 10 for LLM Applications
- MITRE ATLAS
- NIST AI RMF
- CSA AI Controls Matrix
Using
- Garak
- PyRIT
- Promptfoo
- Burp Suite
- ModelScan
Before you launch
Test Your AI Feature the Way an Attacker Would
A one- to two-week review of your chatbot, copilot or agent, with every finding reproduced, fixed and retested. NDA signed before we touch anything.
Book a Security Review
What Clients Say About Working With Us
“I had the pleasure of working with Sajal Tech on a travel website, and it was an incredible experience from start to finish. They demonstrated professionalism, clear understanding of requirements, and excellent communication throughout.”
How We Run Security Engagements
Security testing touches sensitive systems. These rules apply to every engagement.
Written rules of engagement
Scope, test windows, allowed techniques and contacts agreed in writing before testing starts, with production testing only if you approve it.
Severity-ranked findings
Each issue is rated by impact and likelihood, mapped to an OWASP LLM category and delivered with steps to reproduce it.
Confidential handling
Test data, findings and credentials are stored encrypted, shared only with named contacts and deleted at the end of the engagement.
Retest included
After your team or ours applies fixes, we retest every high and critical finding and issue an updated report.
Security Tools and References We Use
We combine open-source scanners with manual testing, and map results to the frameworks your security team already knows.
- OWASP Top 10 for LLM Applications
- MITRE ATLAS
- NIST AI RMF
- CSA AI Controls Matrix
AI Security by Industry
The attack paths are similar everywhere. What leaks, and what it costs when it does, depends on the sector.
Assistants that touch accounts or payments are tested for data exposure and tool abuse before any customer sees them.
- Account data leakage tests
- Payment action confirmation
- Fraud prompt red teaming
- Regulator-ready evidence
FinTech
What AI Security Work Costs
Indicative starting prices. You get a fixed quote once scope, tools and environments are agreed.
LLM security review
Threat model and focused test of one AI feature against the OWASP LLM Top 10.
From
$2,500
1 to 2 weeks
- Threat model
- Injection and leakage testing
- Severity-ranked report
- One retest
Agent security assessment
Full test of an agent or copilot with its tools, retrieval and integrations.
From
$7,500
2 to 4 weeks
- Tool and permission review
- RAG access control testing
- Supply chain review
- Guardrail recommendations
Secure AI program
Guardrails, monitoring and recurring red teaming across your AI portfolio.
From
$20,000
2 to 4 months
- AI gateway and guardrails
- CI attack test suite
- SIEM alerts and playbooks
- Quarterly retesting
How an AI Security Engagement Runs
Most assessments take one to three weeks, depending on the number of AI features, tools and data sources in scope.
2 to 3 days
Threat modeling
Map data flows, trust boundaries, tools and users, and decide which attack scenarios matter most for your product.
- Threat model
- Rules of engagement
1 to 2 weeks
Testing
Automated attack runs plus manual testing of injection, leakage, tool abuse, supply chain and abuse cases.
- Findings log
- Proof of concept attacks
1 to 3 weeks
Remediation
Fix guidance for your engineers, or we implement guardrails, permission changes and sandboxing directly.
- Fixes merged
- Guardrail config
- Updated tests
3 to 5 days
Retest and report
Verify fixes, add attack cases to your CI so they stay fixed, and deliver an executive and technical report.
- Final report
- CI security tests
AI Security Questions
Still deciding? A 30-minute call usually answers the rest.
Ask Us DirectlyNot with current models. No filter catches every injection, so we reduce both the chance and the impact: isolate untrusted content, detect known patterns, and limit what the model's tools can do so a successful injection causes little harm.
A normal pen test covers your web app, APIs and infrastructure. AI testing adds attacks through natural language, documents and tool outputs, and checks what the model can be tricked into revealing or doing. We recommend both, and can coordinate with your existing pen test provider.
It is a community list of the most common risks in LLM apps, including prompt injection, sensitive information disclosure, supply chain issues, excessive agency and unbounded consumption. We use it as the baseline for test coverage and reporting.
The provider secures the model and its platform. You are responsible for your prompts, data, tools, permissions and how outputs are used. Most real incidents happen in that application layer, which is what we test.
Yes, mostly through design. Use zero-retention enterprise endpoints, filter retrieval by the user's permissions, redact PII before it reaches the model and scan outputs. We also check that system prompts contain nothing you would not want exposed.
Only with written approval and agreed test windows. We prefer a staging copy with realistic data and the same tools, and keep production tests to safe, non-destructive checks.
An executive summary, a technical report with each finding, its severity, reproduction steps and fix, plus attack test cases you can add to CI. After fixes we retest and update the report.
A focused review of one AI feature starts from $2,500. A full agent assessment with tools and retrieval starts from $7,500, and ongoing programs with guardrails and recurring tests start from $20,000.
Let's build your next product together
Book a free strategy call and leave with a clear plan and estimate. No commitment.
- 01Pick a time that suits you
- 0230 minutes on scope, stack, timeline and budget
- 03Fixed-price proposal, NDA on request