BUILDING AI AGENTS FOR CYBER DEFENCE

Threats move in seconds.
We are building agents
that move faster.

Project Kavach is building AI agents that detect phishing attempts, halt ransomware before it spreads, and close vulnerabilities from prioritisation through to verified remediation.

Get in touch Our AI agents
44% of all confirmed breaches involved ransomware in 2025 Verizon DBIR 2025
20% of breaches started with vulnerability exploitation, up 34 percent year over year Verizon DBIR 2025
$4.88M average global cost of a single data breach IBM Cost of a Breach 2024
the threat landscape

The data that defines
what we are building to solve

The 2025 Verizon Data Breach Investigations Report analysed over 22,000 real-world security incidents. Every design decision at Project Kavach traces back to what that data says.

44%

Ransomware is present in nearly half of all confirmed breaches globally

The 2025 Verizon DBIR, which analysed 22,052 incidents across 139 countries, found ransomware in 44 percent of all confirmed breaches, up from 23 percent in 2024. Critically, 88 percent of ransomware incidents targeted small and mid-sized organisations. Ransom groups no longer discriminate by size; they scale demands proportionally to revenue. The operational damage, recovery time, and reputational cost far exceed the ransom figure itself.

Verizon DBIR 2025

Phishing is the primary delivery mechanism for credential theft and ransomware

IBM X-Force 2025 recorded an 84 percent year-over-year surge in phishing emails carrying infostealer malware, with early 2025 data suggesting a 180 percent increase against 2023 baselines. Attackers now use AI to generate contextually convincing lures at scale. The median time from email open to credential submission is under 60 seconds, faster than any human review process can operate.

IBM X-Force Threat Intelligence Index 2025

Vulnerability exploitation is outpacing patching cycles at every organisation

Vulnerability exploitation as an initial access vector rose 34 percent year over year in the 2025 DBIR, now accounting for one in five breaches. For critical edge device vulnerabilities, the median time between public disclosure and active mass exploitation was zero days. IBM X-Force found that 60 percent of the top CVEs discussed on dark web forums had weaponised exploit code within two weeks of disclosure.

Verizon DBIR 2025
project kavach ai agents

Three AI agents. Each built
for a specific attack category.

Project Kavach is building three purpose-built AI agents. Each has deep, focused capability in one domain. Together they form a coordinated defence layer across phishing, ransomware, and vulnerability management.

Phishing AI Agent
Lens
Email Threat Intelligence Agent

Lens analyses every inbound message using natural language understanding, sender behaviour modelling, link graph analysis, and attachment inspection calibrated to your organisation's communication patterns. It identifies phishing, business email compromise, AI-crafted impersonation, and credential harvesting attempts before any user interacts with the content.

  • Full header analysis and sender reputation scoring against live threat intelligence
  • NLP intent detection for business email compromise, pretexting, and spear-phishing
  • Automated quarantine with analyst notification and complete incident documentation
  • Learns per-organisation communication baselines to reduce false positives continuously
IBM X-Force 2025: 84% surge in infostealer phishing
Ransomware AI Agent
Brace
Ransomware Detection and Containment Agent

Brace monitors process behaviour, file system activity, memory access patterns, and lateral movement signals continuously. The moment ransomware behavioural indicators appear — mass file modification, shadow copy deletion, abnormal privilege escalation — Brace isolates the affected system and stops propagation across the network in seconds.

  • Behavioural detection independent of signatures, effective against novel ransomware families
  • Automated endpoint isolation to block lateral movement before spread occurs
  • Forensic evidence preservation for post-incident root cause analysis
  • Coordinates with Lens when phishing is the ransomware delivery mechanism
Verizon DBIR 2025: ransomware in 44% of all breaches
Vulnerability Management AI Agent
Seal
Vulnerability Prioritisation and Remediation Agent

Seal works downstream of your existing scan tools. It ingests raw vulnerability findings and applies intelligence to them: ranking every finding by real-world exploitability, active threat actor usage, and business impact. It then drives the full remediation lifecycle — from prioritised action plan through to verified closure.

  • Ingests findings from existing scanners, applies risk ranking based on exploitability and asset criticality
  • Automated patch application where feasible, with human confirmation required before execution on critical systems
  • Where no patch exists, Seal generates a concrete mitigation plan with specific compensating controls
  • Tracks every finding through to verified closure and re-confirms each remediation is complete
Verizon DBIR 2025: vulnerability exploitation up 34% YoY
how it works

Designed around how
breaches actually unfold

Each Project Kavach AI agent operates on the same four-phase loop, aligned with the NIST incident response lifecycle and grounded in MITRE ATT&CK adversary behaviour mapping.

01

Connect without replacing your stack

Project Kavach integrates with the security tooling you already operate. The AI agent layer sits above your existing infrastructure, consuming and correlating telemetry without requiring a replacement cycle.

02

Detect through behaviour, not just signatures

Agents correlate signals across multiple sources simultaneously. A single event that appears benign in isolation becomes significant when placed in context with other activity. That correlation runs continuously at machine speed.

03

Act or escalate with full context already prepared

High-confidence findings trigger autonomous action: quarantining messages, isolating endpoints, initiating patch workflows. Ambiguous findings are escalated with the full investigation assembled — no analyst starts from scratch.

04

Verify closure and improve over time

Every resolved finding is confirmed closed, not just actioned. Each case feeds back into the agent model of your environment, improving precision and reducing false positives continuously.

EVIDENCE BASIS FOR THIS APPROACH
Zero days
median time between critical vulnerability disclosure and active mass exploitation for edge devices in 2025, making calendar-based patching cycles obsolete
Under 60 sec
median time for a user to click a phishing link after opening the email, faster than any manual alert triage process can respond
$2.2M
average reduction in breach cost when AI is used extensively in prevention workflows versus organisations with no AI in security operations
100 days
faster breach identification and containment for organisations that deploy AI and automation across their security operations centre
design principles

How Project Kavach thinks
about building AI agents

Speed as a hard constraint

When phishing clicks happen in under 60 seconds and ransomware encrypts thousands of files per minute, response time is not a preference. Every architectural choice is measured against time to action first.

🔬

Built on published research

Threat models are derived from the Verizon DBIR, IBM X-Force, MITRE ATT&CK, and CISA Known Exploited Vulnerabilities. We build on what has been observed and documented, not hypothesised.

🎯

Fewer alerts, higher confidence

Alert fatigue is a documented cause of security programme failure. The goal is not more visibility — it is fewer, higher-confidence findings that analysts can act on without hesitation.

🧑

Human decision at the boundary

Automation handles speed and scale. Human judgement handles consequence. On Seal's patch actions and Brace's isolation decisions, a human approves anything with irreversible business impact.

🔗

Three agents, one kill chain

Lens, Brace, and Seal are designed to work independently and as a coordinated system. A phishing-delivered ransomware attack that exploits an unpatched vulnerability is covered end to end.

📡

Intelligence that improves continuously

Each AI agent ingests live threat intelligence and learns from your specific environment. Detection does not stay static as attackers adapt — it improves with every resolved incident.

market and research basis

Why AI is the only credible
answer to the current threat

All figures are drawn from primary research published by Verizon, IBM, and independent market research firms. Each is linked directly to its source document.

AI IN CYBERSECURITY — PROJECTED MARKET SIZE (USD)
2022approx. $15B
2024$25.4B
2028 (projected)approx. $60B
2030 (projected)$93.8B
Grand View Research: AI in Cybersecurity Market Report 2025
Polaris Market Research: AI in Cybersecurity Market 2025
24.4%
compound annual growth rate of the AI cybersecurity market through 2030
30%
of all intrusions in 2024 began with valid compromised credentials
70%
of all attacks IBM X-Force responded to targeted critical infrastructure, with over a quarter caused by vulnerability exploitation
88%
of ransomware incidents in 2025 targeted small and mid-sized organisations
in development

The work is underway.
The mission is clear.

Project Kavach is actively building AI agents for cyber defence. If you want to follow what we are building or explore a conversation, we would like to hear from you.

Get in touch
about project kavach

We are building AI agents because defenders need tools that match the speed of the threat

Project Kavach takes its name from the Sanskrit word for armour not a barrier you hide behind, but protection that moves with you. That distinction shapes how we are building every AI agent in the platform.

The 2025 Verizon Data Breach Investigations Report analysed over 22,000 real-world incidents and found ransomware in 44 percent of breaches, vulnerability exploitation growing 34 percent year over year, and phishing still the dominant delivery mechanism for both. The IBM X-Force 2025 Threat Intelligence Index found that attackers now deploy infostealer malware via phishing at 84 percent higher weekly volumes than the prior year. These are not incremental shifts. The threat has fundamentally changed in speed and sophistication.

We are building Project Kavach because the gap between how quickly attacks move and how quickly defences respond cannot be closed by human-only processes. Lens, Brace, and Seal are each built for one of the three most consequential and well-documented attack categories, with the depth and specificity that a general-purpose platform cannot achieve.

Every design decision traces back to published research. Every AI agent action will be logged, explainable, and auditable. We build on what is observed, not assumed.

research grounded

Threat models are derived from the Verizon DBIR, IBM X-Force, MITRE ATT&CK, and CISA guidance. We do not build on intuition when data is available.

explainable by design

Every action an AI agent takes will be logged, traceable, and explainable. Security tools must earn trust. We will not build systems that cannot account for their decisions.

depth over breadth

Three AI agents built with real depth are more valuable than ten shallow features. Project Kavach is narrow by intention and rigorous within that focus.

human judgement at the boundary

Automation handles speed and scale. Human analysts handle consequence and context. On the actions that matter most, a human confirms before execution.

"To close the speed gap between how fast attackers move and how fast defenders can respond — by building AI agents that are precise, autonomous, and always on."
primary research references
Verizon (2025). 2025 Data Breach Investigations Report. verizon.com/business/resources/reports/dbir/
IBM Security and Ponemon Institute (2024). Cost of a Data Breach Report 2024. ibm.com/reports/data-breach
IBM X-Force (2025). X-Force Threat Intelligence Index 2025. ibm.com/thought-leadership/institute-business-value/report/2025-threat-intelligence-index
MITRE Corporation. ATT&CK Enterprise Matrix. attack.mitre.org
CISA. Known Exploited Vulnerabilities Catalogue. cisa.gov/known-exploited-vulnerabilities-catalog
contact

Get in touch

Project Kavach is building AI agents for cyber defence. If you want to follow our progress, explore a collaboration, or talk about where this space is heading, we would like to hear from you.

We are building something that matters

Project Kavach is actively building three AI agents. Lens for phishing, Brace for ransomware, and Seal for vulnerability management. We are always open to speaking with researchers, security professionals, and anyone serious about the future of cyber defence.

IN ACTIVE DEVELOPMENT

Project Kavach is actively building its AI agents. If you are a researcher, security professional, or potential collaborator, send a message and we will get back to you.

Send us a message

We respond to every message personally.
Message received

We will be in touch soon.