01
Understand AI
Study how AI learns, reasons and acts.
We research how AI works, secure how it acts, and build systems that enterprises can rely on.
Programs

Engagements


Research





Why Cosmipher
The same understanding that helps us find weaknesses also helps us build better AI systems.
01
Study how AI learns, reasons and acts.
02
Protect every path an agent can take.
03
Put the same foundations to work.
Engagements and programs
Programs and engagements show how we apply AI security and governance to real workflows.
EngagementIndiaMART
EngagementAngel One
ProgramNVIDIA Inception
Enterprise engagement
We work with IndiaMART to secure its agentic systems, assess exposure and design controls for agent actions.

Enterprise engagement
We built AdReview to screen advertisements against compliance requirements and route evidence to reviewers.

Program
Cosmipher is part of NVIDIA's program for AI startups building at the frontier.
Program
A member of our team has Daybreak Blue access for responsible cybersecurity research.
Program
Our agents run on Google infrastructure. Our researchers responsibly disclose AI vulnerabilities to Google.
Freedom v1.1
Freedom v1.1 is built in India for cybersecurity research, authorized testing and enterprise defense and offense.
602B
Total logical parameters
Gold-standard
Trained on curated cybersecurity big data
Mixture of experts
Active parameters: 8B during prefill, 16B during decoding
Multimodal image and text input
Input
Freedom v1.1 beside its strongest peer on four cybersecurity benchmarks.
CyberGym leaderboard / one-trial model track
QCRI audit / 2026
QCRI audit / 2026
QCRI audit / 2026
97.00%
20-case Security Showcase
92.34%
431-scenario full catalog. 431 demanding scenarios across 51 cyber domains.
~400 TB
Compressed cybersecurity knowledge. All five Vs of big data.
Mixture of experts8B during prefill, 16B during decoding
Research
A shared coordinator routes work to the agent harness best equipped for the task.
Intent, risk, permissions and promotion rights are fixed before any work begins.
Agents work under that contract, and every trajectory is recorded with its evidence.
Independent assurance judges each outcome. Model confidence never stands in for evidence.
The Experience Compiler promotes repeated verified reasoning into the lightest reusable form: an episode, a procedure, a skill or a deterministic tool.
Useful successes and failures keep their provenance, so the system never starts from zero.
Repeated verified reasoning migrates to cheaper reusable capability once held-out evidence supports it.
Every active artifact stays scoped, provenance-linked and revocable. Contradicted evidence re-evaluates everything built on it.
IIT Madras
The University of Texas at Austin
IIM Ahmedabad
IIM Raipur
IIT KanpurMeasured across our CSI test runs.
80%
Token-cost reduction
40×
Speed-up
1,000s
Parallel agents
Secure it
A model can retrieve data, call tools, delegate tasks and change business records. Each connection adds a security boundary.
The wider catalog expands these roles into 32 tools and modules across the two groups below.
01
Discover models, agents, identities, tools and data connections.
Explore on the AI Security site02
Evaluate interactions and tool actions at supported boundaries, then apply the policy and retain a reviewable decision.
Explore on the AI Security site03
Exercise authorized failure paths, preserve reproducible findings, guide remediation and replay the original attack after a change.
Explore on the AI Security site04
Connect exact artifact identity, composition, lineage, inspection, ownership and release approval.
Explore on the AI Security siteDirection
Secure the age of intelligence by helping organizations deploy and operate AI they can rely on.
Build AI in India that enterprises worldwide can trust with their most demanding work.
Use the model, the cybersecurity corpus and the capability atlas to test security capability with stronger evidence.
Connect discovery, testing, runtime policy and release evidence across more enterprise systems.
Carry shared context, permissions and verified experience into the applications we build. Use what we learn there to improve the underlying systems.
We measure progress by what teams can use securely.
Ciphering the cosmos means understanding how models, agents and systems work so we can secure them and build AI people can rely on.