The real attack path stays out of view
Individual checks pass while the chain between identity, data, and authority remains untested.
Security audits & AI red teaming
Somveer Cyber Solutions
India — Working globally / Est. 2026
We look beyond the surface.We test the boundariesthat keep systems safe.
Bring us the system and the questions you cannot answer yet.
AI red teaming, application security, identity testing, and evidence-led validation.
Connecting adversarial thinking with evidence, applications with AI, findings with fixes — security research that follows the whole chain. ／
01 — IssueThe faster systems evolve,
The faster systems evolve,
the more their trust boundaries matter.
AI has made software faster to build and easier to connect.
Every new connection brings a new assumption worth testing.
Permissions drift across connected services
A valid session reaches a capability that the original trust boundary should never have granted.
Agents inherit more authority than they need
Untrusted input can influence how a model uses tools, retrieves data, or takes actions.
Scanner output gets mistaken for impact
A severity label arrives before the finding has been reproduced and challenged with controls.
The report ends before the fix is testable
Engineering teams need a clear root cause and a way to verify that the boundary holds.
More automation creates more uncertainty
Fast testing is useful when scope, evidence, and human judgment remain part of the process.
02 — ApproachFollow the whole chain.
Follow the whole chain.
Prove the boundary that failed.
We connect manual investigation, automated testing, and disciplined validation to understand how a system actually behaves.
Scope, hypothesis, evidence, and remediation form one continuous process. The result is a finding an engineer can reproduce, assess, and fix.
SCOPEINVESTIGATEVALIDATEDELIVERTHINK → TEST → LEARN ↗
03 — BlueprintResearch that gets sharper
Research that gets sharper
with every iteration.
How human judgment, automated testing, and evidence
come together in our research workflow.
aisafesafe.me / RESEARCH BLUEPRINT 01
The independent security research loop
A question becomes evidence, then a fix and a repeatable check.
01 — Participants
RESEARCHERInvestigates and validates
PRODUCT TEAMContext, scope, and remediation
USERThe boundary we protect
02 — Operating & learning loop
AUDIT TOOLINGScoped testingHUMAN REVIEWEvidence and judgment
01 / QUESTIONDefine the hypothesis02 / TESTInvestigate behavior03 / EVIDENCEChallenge with controls04 / DELIVERYExplain and remediate05 / LEARNINGKeep the regression
03 — Accumulation
Knowledge compounds.
Each investigation improves the questions, checks, and judgment used in the next one.
Test 1Test 2Test 3
04 — ValueAutomation and human judgment,
Automation and human judgment,
working in one process.
Five things a useful security engagement should leave behind.
A precise security question
Turn a broad concern into a testable hypothesis, grounded in your product, assets, and threat model.
Reproducible evidence
Record the behavior, relevant controls, and impact chain so the finding stands up to technical review.
Calibrated impact
Separate observed outcomes from simulations, assumptions, and conditions that still need validation.
Actionable remediation
Explain which boundary failed and what an engineering team can change to enforce it.
A way to verify the fix
Define the controls and regression checks that should pass after the issue is addressed.
Stance
How we work with AI
AI helps explore code and automate repetitive work. A researcher remains responsible for scope, validation, severity, and every submission.
05 — ServicesFrom an open question
From an open question
to a defensible finding.
Choose the work your product needs.
We agree on scope, permissions, and deliverables before testing.
01AI & Agent
AI & Agent
Red Teaming
Test the boundaries around models, tools, and data
Scoped adversarial evaluation for LLM applications and agents. Examine how untrusted input affects tool use, retrieval, sensitive data, and the handoff between model output and product behavior.
02Application &
Application &
API Audits
Follow identity and authorization through the whole system
Scoped security audits of web applications, APIs, authentication, authorization, and business logic. Trace connected weaknesses and demonstrate the minimum evidence needed to establish realistic impact.
03Security
Security
Automation
Build repeatable testing without losing human judgment
Custom Python tooling for scoped discovery, authenticated testing, evidence collection, and validation. Repeatable checks support security audits, with human review of findings and impact.
06 — ResearchThe evidence behind
The evidence behind
the practice.
Selected investigations by Somveer Yadav.
Each finding is shown with its actual reported status.
FEATURED / OAUTH & EMBEDDED WALLETSWhen a login redirect
When a login redirect
becomes wallet control.
A research report tracing an OAuth destination boundary into sessions, embedded-wallet authority, and persistent account access.
01Login destinationFirst trust boundary
↓02Account sessionDownstream identity
↓03Wallet authorityReported impact
One request.
500 resolver executions.
GraphQL alias amplification bypassed a query-complexity limit. Reported to white.market; the fix was deployed.
Cloud credentials.
No authentication.
An unauthenticated upload endpoint issued temporary AWS credentials with unintended access capabilities.
A model artifact.
A code-execution boundary.
Research into unsafe pickle deserialization in the SageMaker Python SDK and its developer-workstation impact.
Researcher-reported disclosure status. Submitted and in-review findings are pending vendor validation. Case summaries distinguish observed behavior, test conditions, and simulated impact. Company and product names identify research subjects, not clients or endorsements.
07 — ProcessAsk precisely. Test carefully.
Ask precisely. Test carefully.
Keep the evidence.
Every engagement follows a clear research process.
- Step 01
Understand
Learn the product, the assets you control, and the security question that matters.
- Step 02
Scope
Agree on permissions, accounts, constraints, testing limits, and deliverables.
- Step 03
Investigate
Follow the behavior with manual research and carefully bounded automation.
- Step 04
Validate
Use negative controls, repeatable evidence, and realistic impact analysis.
- Step 05
Deliver
Explain the root cause and define remediation and regression checks.
How we run engagements
- Agree on written scope before testing
- Prove the impact with the minimum necessary evidence
- Use automation to support careful investigation
- Keep human review at every submission decision
- Share reasoning, limitations, and a path to remediation
08 — DifferenceA finding should stand up
A finding should stand up
to the next question.
What evidence-led research adds to a security assessment.
The questionWhat the research should establish
“The endpoint responded.”Did it cross a real authorization boundary?
“The scanner says critical.”What impact was actually reproduced?
“The happy path works.”Do negative controls reject invalid requests?
“The exploit is possible.”Which conditions and user actions are required?
“The issue is fixed.”Which regression check demonstrates the fix?
09 — What staysMore than a finding.
More than a finding.
A better way to test the next one.
Clear evidence and reusable knowledge should outlast an engagement.
Reproduction stepsNegative controlsRoot-cause analysisScope documentationImpact boundaries
Remediation guidanceRegression checksEvidence packagesResearch toolingTechnical handover
Look beyond the surface.
Follow the system.
Let the evidence decide.
AI can produce thousands of hypotheses.
Research asks which one the evidence supports.
A useful investigation follows a weakness beyond its first symptom, then challenges its own conclusions.
We distinguish a live observation from a simulation, a submitted claim from a confirmed finding, and an apparent impact from one we can defend.
That discipline is the foundation of this practice.
Scope
Know the boundaries of the engagement
Evidence
Make every claim reproducible
Judgment
Calibrate the actual impact
Clarity
Explain the root cause and the fix
Learning
Turn the result into a reusable check
11 — The practiceSecurity audits.
Security audits.
Accountable by name.
aisafesafe.me is the security audits and research practice of Somveer Cyber Solutions,
a Udyam-registered micro enterprise based in India.
SY / INDEPENDENT RESEARCHER
Somveer Yadav
Founder · Security ResearcherIndependent security research since August 2024. Focused on authentication, authorization, APIs, and evidence-led vulnerability research.
Diploma in Cyber Security Engineering. Google Cybersecurity Professional Certificate. Selected for the sponsored C|PENT AI practical exam; in progress.
SECURITY AUDITS / CLEAR DELIVERABLES
Scoped security audits
Applications · APIs · IdentityEvery engagement starts with written authorization and an agreed scope. Testing follows defined assets, accounts, permitted methods, and stop conditions.
01Agree scope
02Document evidence
03Retest fixes
Registered enterpriseSomveer Cyber SolutionsUDYAM-HR-15-0041026Registered 09 Aug 2026 · Haryana, India
12 — FAQA few questions
A few questions
before we start.
Get in touch even if your scope is still taking shape.
Can we talk before the scope is fully defined?+
Yes. Start with your product, the assets you control, and the questions you want answered. We can then agree on written scope, testing permissions, deliverables, and commercial terms.
Do you work with AI labs and product teams?+
The practice welcomes scoped evaluation and research opportunities with AI labs and product teams, including products built on Claude, GPT, and other models. No affiliation with Anthropic or OpenAI is claimed.
Can we request just an application or API assessment?+
Yes. Services can be scoped independently. The assessment can focus on a specific application, identity flow, API, authorization boundary, or business-logic concern.
How is sensitive research handled?+
Testing follows written authorization and program scope. Detailed evidence is shared through the agreed channel, and publication follows coordinated disclosure. These website summaries omit operational secrets and runnable exploit instructions.
What do the research statuses mean?+
Resolved, accepted, submitted, and in review represent different stages. A submitted or in-review report is not presented as vendor-confirmed. Statuses reflect the researcher’s documented report history; the individual case summaries explain their sources and limitations.
How are pricing and timelines decided?+
They depend on scope, complexity, access, and the depth of testing. Share your goals and preferred timeline so that scope and commercial terms can be discussed before work begins.
What should we prepare for the first conversation?+
A short brief covering your product, assets you control, desired outcomes, and preferred timeline. Avoid sending credentials or sensitive evidence in the first message.
Let’s talk
Contact us
Tell us what you are building and what you want to understand.