Ares V1.0 · Dagger
Every weight built to breach.
Dagger is Assail’s proprietary 14-billion-parameter model, trained from the ground up for one mission: finding and exploiting vulnerabilities across APIs, web applications, and mobile applications. Not a chatbot with a Burp Suite plugin: an offensive security model where every parameter exists to think like an attacker.
The Ares Proprietary Model
Purpose-built to find weaknesses, chain exploits, and validate impact.
Frontier models were trained to be helpful, harmless, and broadly capable. Dagger was trained for a single job, and it does that job at machine speed, at machine cost, with zero hedging.
Purpose-Built
A 14-billion-parameter offensive security model architected by a 26-year hacker and trained on real exploit chains, not CVE summaries, not scanner output.
Self-Validating
Validation is a first-class capability baked into the model. Every finding is confirmed exploitable before it reaches a customer report.
Self-Sharpening
Continuous training promotes a sharper model into production on a weekly cadence, with no upgrades to buy and no releases to wait for.
Trained for the Breach
No refusals. No moralizing. No system-prompt jailbreaks holding a safety layer at bay. Every parameter exists to think like an attacker.
The Wrapper Problem
Frontier models were built to help. Dagger was built to breach.
A general-purpose training objective is fundamentally at odds with offensive security work. Here is what changes when you run a purpose-built model instead of a frontier wrapper.
Dimension
Wrapped Frontier Model
Dagger
Training objective
Helpful, harmless, broadly capable: refuses dangerous requests.
Trained to find weaknesses, chain exploits, synthesize payloads, and validate impact.
Offensive tasks
Hedges, refuses, and moralizes mid-engagement.
Zero refusals. Zero moralizing. Zero wasted cycles.
Economics
Per-token billing at frontier prices: the math breaks at scale.
Runs on our infrastructure, at our cost basis, with no per-token markup.
Validation
Flags everything and calls it a finding.
Confirms every finding is exploitable before it reaches a report.
Control
Depends on an upstream provider’s pricing, policies, and outages.
Runs in infrastructure we own and control: no upstream provider.
Signal, Not Noise
A model that knows the difference between theory and exploit.
A SQL injection on an internal admin login is critical. The same injection on a marketing page behind three layers of WAF, with zero data access, may be unexploitable noise. Generic models cannot tell the difference: they flag everything and call it a finding. Dagger was trained with security-adapted reward formulations that account for the context-dependent nature of real exploitation.
The Validator Agents
Themis
Validator agent
Aletheia
Validator agent
Argus
Validator agent
Mnemosyne
Validator agent
Each calls Dagger to confirm exploitability
Dagger validates the finding against real impact
Only findings confirmed exploitable reach the customer report. False positives are not tolerated: they are eliminated at the model level.
Continuous Training
Weekly improvements. A production-grade pipeline named Javelin.
Dagger is not a static asset that ages between releases. Every week, an automated pipeline retrains, benchmarks, and (only if it clears the gate) promotes a sharper model into production.
01
Export
Every Sunday at 02:00 UTC, Javelin’s automated training pipeline exports new synthetic engagement data.
02
Threshold check
The pipeline confirms a minimum of 100 new training examples before it commits to a run.
03
Fine-tune
The inference cluster scales down. QLoRA fine-tuning runs at rank 16, alpha 32, learning rate 5e-5, for one epoch.
04
Merge
The new adapters merge into the base model and the inference cluster scales back up.
05
Benchmark gate
The gate enforces a minimum score of 0.80 and rejects any regression greater than 5%.
06
Promote or roll back
Clear the gate, and the checkpoint promotes to production. Fail, and it rolls back automatically.
No upgrades to buy. No releases to wait for. Customers wake up Monday with a sharper Dagger under every agent in their tenancy.
Published Research
Documented architecture, not marketing vapor.
The co-evolutionary architecture powering Dagger is documented in a peer-reviewed framework paper authored by Assail’s research team. These are architectural decisions, made transparent so the offensive security industry can evaluate the technology it depends on.
Group Relative Policy Optimization
The reinforcement-learning method used to train the model.
Uncertainty-based frontier filtering
How the model selects where to focus its learning.
Security-adapted reward formulations
Rewards tuned for the context-dependent nature of real exploitation.
Adversary-Breacher dynamics
The co-evolutionary dynamic at the core of the architecture.
What You Actually Get
Ask a vendor one question: what model are you actually running?
If the answer is GPT, Claude, or Gemini with a custom prompt, you are paying enterprise security pricing for consumer infrastructure. Here is what changes when you choose a proprietary, purpose-built model.
Economics that scale with your attack surface
A single enterprise API engagement can generate millions of inference calls: recon, fingerprinting, payload synthesis, validation, reporting. At frontier per-token pricing, the math breaks before the first scan finishes. Dagger runs on our infrastructure, at our cost basis, with zero per-token markup passed to you.
Zero refusals. Zero moralizing. Zero wasted cycles.
Vendors wrapping frontier models rely on system prompts and jailbreaks that break every time the upstream safety layer updates. Dagger will fingerprint your stack, enumerate endpoints, chain vulnerabilities, and synthesize working exploit payloads without a single refusal or hedge because it was never trained to.
Sovereignty over your security stack
Depend on a frontier wrapper and you depend on every upstream decision: pricing changes, deprecation timelines, regional availability, policy revisions, and cascading outages. Dagger runs in infrastructure we own and control. No upstream provider. No third-party rate limit. No consumer-grade safety filter between you and the work.
Built by someone who has done the work
Dagger was architected by an offensive security practitioner with 26 years in the field, two prior company exits, research cited in U.S. Congressional proceedings, a published book on adversarial security (Hacking Connected Cars, Wiley & Sons), and hacking tools exhibited at The Mob Museum. Wrapper vendors hire prompt engineers. Assail hires hackers.
FAQs
Questions about the model? Answers.
The most common questions about Dagger, Javelin, and how the model is built.
Start today
Get full access → $249