Khasm · PROGRAM ROADMAP
A phased, fully-funded program from core attack library to fleet-wide AI coverage
Khasm is scoped as a multi-phase build, not a single release. Each phase below carries its own investment, team, budget breakdown, and exit-criteria milestones, sequenced so that later phases draw directly on the attack-library and compliance work completed earlier — no phase starts until its dependencies are met.
Core attack library & findings register
PlannedPhase 1 — core platform buildout
Resources
Build the initial prompt-injection and jailbreak attack suites against representative LLM architectures.
Build the campaign runner, findings register, and transcript-capture pipeline.
Aligns reporting to OWASP LLM Top 10, SAMA FEER, and NCA ECC-2 patterns.
Owns cross-workstream sequencing, reporting, and design-partner readiness.
Budget breakdown
Attack-library engine, campaign runner, findings register, and transcript-capture pipeline.
OWASP LLM Top 10, SAMA FEER, and NCA ECC-2-aligned reporting formats and documentation.
Onboarding tooling and integration test rigs against representative LLM copilots and agents.
Reserve against attack-library and integration-timeline risk.
Plan of action
- 1Build the initial prompt-injection & jailbreak attack suiteAttack-library engineering · 6 weeks
Cover the OWASP LLM Top 10 bypass categories against representative LLM architectures.
- 2Build the campaign runner and findings registerPlatform engineering · 8 weeks
Every successful bypass captured as a reproducible transcript, not a severity score alone.
- 3Align reporting to OWASP LLM Top 10, SAMA FEER, and NCA ECC-2Compliance lead · 6 weeks
Produce report formats regulators and design partners already recognize.
- 4Sign first 2-3 design-partner letters of intentProgram manager · 10 weeks
Confirm the LLM copilots and agents the first live campaigns will run against.
Activities
Prompt-injection & jailbreak attack library
PlannedContinuously updated adversarial prompt suites testing the latest bypass techniques.
Owner Attack-library engineering
Findings register with transcript proof
PlannedEvery successful attack captured as a reproducible transcript.
Owner Platform engineering
Depends on Prompt-injection & jailbreak attack library
OWASP LLM Top 10, SAMA FEER & NCA ECC-2 alignment
PlannedReporting formats and evidence trail aligned to recognized frameworks.
Owner Compliance lead
Sovereign hosting for attack transcripts
PlannedAttack transcripts and model outputs hosted and retained inside the Kingdom by default.
Owner Platform engineering
Depends on Findings register with transcript proof
Exit-criteria milestones
- ✓Prompt-injection and jailbreak attack suite validated against representative LLM architectures
- ✓Campaign runner and findings register live with transcript capture
- ✓OWASP LLM Top 10, SAMA FEER, and NCA ECC-2-aligned reporting formats approved
- ✓2-3 design-partner letters of intent signed
Pilot deployment & hardening
PlannedPhase 2 — pilot deployment & hardening
Depends on Core attack library & findings register
Resources
Build and validate tool-misuse attack campaigns against agents with real tool/function access.
Connect design-partner LLM copilots and agents into continuous attack campaigns.
Owns pilot success criteria and design-partner relationships end to end.
Continues cross-workstream sequencing and pilot reporting.
Budget breakdown
Tool-misuse attack campaigns and data-exfiltration probes for agentic systems.
Connecting production LLM copilots and agents into continuous campaigns across pilot sites.
Pilot-cohort account management and findings-remediation support.
Reserve against integration and remediation-timeline risk across the pilot cohort.
Plan of action
- 1Deploy continuous campaigns against the first design-partner systemIntegration engineering · 4 weeks
Connect the first production LLM copilot or agent and go live with continuous attack campaigns.
- 2Ship agentic tool-misuse and data-leakage attack campaignsAgentic security engineering · 10 weeks
Extend the attack library from text-only jailbreaks to tool-calling agents and data-exfiltration probes.
- 3Harden remediation SLA and automatic re-test workflowPlatform engineering · 6 weeks
A fix ships, a re-test runs automatically, and the finding only closes once it passes.
- 4Run a structured pilot review cyclePilot lead · 12 weeks
Track finding volume, time-to-fix, and re-test turnaround against baseline.
Activities
Agentic tool-misuse testing
PlannedTests whether an agent with tool or function-calling access can be coaxed into an unauthorized action.
Owner Agentic security engineering
Depends on Prompt-injection & jailbreak attack library
Data-leakage & exfiltration probes
PlannedTests whether a system can be tricked into revealing its system prompt or another user’s data.
Owner Agentic security engineering
Depends on Prompt-injection & jailbreak attack library
Remediation SLAs & automatic re-test
PlannedCritical and high findings carry a fix-by clock, with re-test scheduled automatically after a fix.
Owner Platform engineering
Depends on Findings register with transcript proof
First design-partner pilots
PlannedContinuous campaigns running against 2-3 design-partner production LLM copilots or agents.
Owner Integration engineering
Depends on OWASP LLM Top 10, SAMA FEER & NCA ECC-2 alignment, Sovereign hosting for attack transcripts
Exit-criteria milestones
- ✓First design-partner LLM system under continuous attack campaigns
- ✓Agentic tool-misuse and data-leakage attack campaigns validated in a live pilot
- ✓Remediation SLA and automatic re-test workflow live
- ✓Pilot performance report validated against baseline
Fleet-wide expansion & certification
PlannedPhase 3 — fleet-wide expansion
Depends on Pilot deployment & hardening
Resources
Build the multi-system fleet view across an organization’s full AI estate.
Track and add emerging jailbreak and tool-misuse techniques to the attack library.
Opens and manages expansion into SAMA-regulated and NCA-regulated enterprise networks.
Owns the ISO 27001 certification audit.
Budget breakdown
Multi-system fleet dashboard and cross-system findings rollup for a full AI estate.
Continuously updated jailbreak and tool-misuse techniques as new bypass methods appear.
Expansion into SAMA-regulated and NCA-regulated enterprise networks beyond the pilot cohort.
ISO 27001 certification audit and infrastructure scaling for a multi-system fleet.
Reserve against certification-timeline and enterprise-procurement risk.
Plan of action
- 1Design and deliver the multi-system fleet viewPlatform engineering · 10 weeks
One rollup view across every registered LLM copilot and agent, for teams running more than one system.
- 2Stand up the attack-library update pipelineAttack-research engineering · 8 weeks
A recurring process for tracking and shipping emerging jailbreak and tool-misuse techniques.
- 3Open first enterprise expansion accountsEnterprise account lead · 14 weeks
Extend the design-partner model into additional SAMA- and NCA-regulated enterprise networks.
- 4Complete ISO 27001 certification auditCompliance lead · 12 weeks
Close out formal certification building on the architecture pre-aligned during phase 1.
Activities
Multi-system fleet view
PlannedA rollup view across every registered LLM copilot and agent for teams running more than one system.
Owner Platform engineering
Depends on First design-partner pilots
Attack-library updates for emerging techniques
PlannedContinuously updated jailbreak and tool-misuse techniques as new bypass methods appear.
Owner Attack-research engineering
Depends on Prompt-injection & jailbreak attack library
Enterprise network expansion
PlannedExtends beyond the pilot cohort into additional SAMA- and NCA-regulated enterprise networks.
Owner Enterprise account lead
Depends on First design-partner pilots
ISO 27001 certification
PlannedComplete formal certification building on the pre-aligned architecture.
Owner Compliance lead
Depends on OWASP LLM Top 10, SAMA FEER & NCA ECC-2 alignment
Exit-criteria milestones
- ✓Multi-system fleet view live across every registered LLM copilot and agent
- ✓Attack-library update pipeline shipping emerging techniques on a recurring cadence
- ✓First enterprise expansion accounts live beyond the pilot cohort
- ✓ISO 27001 certification achieved
Make AI red-teaming a habit, not an audit
We're selecting design-partner teams running LLM copilots or agentic AI in production for the first pilot cohort. Book a briefing to see the product and discuss fit.