An abliterated (refusal-suppressed) build of fdtn-ai/Foundation-Sec-8B-Instruct, Cisco Foundation AI's instruction-tuned, cybersecurity-specialized 8B model.
This is the Instruct variant, not the base model — it follows chat-style instructions out of the box. Refusal behavior was reduced using Heretic, an open-source directional-abliteration tool, so the model engages with legitimate security-research prompts (vulnerability analysis, exploitation theory, red-team planning, payload crafting) that an over-aligned general model often declines.
⚠️ Decensored model. Safety filtering has been deliberately reduced. Outputs are unfiltered and may be wrong or harmful. You are responsible for how you use it and for verifying every output. Intended for research, testing, and controlled environments. See Responsible use.
Why this base
Generic abliterated chat models lose the plot on security specifics. Foundation-Sec-8B-Instruct is continued-pretrained on a cybersecurity corpus (threat-intel reports, vulnerability databases, incident-response docs, standards) and reports cyber-threat-intelligence performance comparable to a 70B general model at 8B. Starting from a domain model means the abliteration only has to remove refusals — the knowledge is already there.
1curl http://127.0.0.1:8089/v1/chat/completions \2 -H "Authorization: Bearer your-local-key" -H "Content-Type: application/json"\3 -d '{"messages":[{"role":"user","content":"Map CVE-2021-44228 to its CWE and explain the root cause."}]}'
Ollama:
ollama run hf.co/e12ex2/Foundation-Sec-8B-Instruct-heretic:Q5_K_M
Chat template: Llama 3.1 (<|begin_of_text|>, <|start_header_id|>…<|end_header_id|>, <|eot_id|>). If you see special tokens leaking into output, your runtime's template or eos_token_id is misconfigured — fix that before trusting results.
Intended use
Built for security practitioners and researchers doing the work mainstream models obstruct: CVE/CWE/CVSS analysis, MITRE ATT&CK mapping, alert triage and SOC summarization, threat modeling, attack-path simulation, OWASP-oriented code review, and red-team planning for authorized engagements.
Limitations
It's 8B. It will miss things a frontier model catches, and it hallucinates. Treat output as a fast first pass, not ground truth.
Knowledge model, not a code-gen powerhouse. Strong at security reasoning and classification; weaker at discovering novel bugs in large unfamiliar codebases or writing complete, correct exploits. For heavy code analysis, pair it with a dedicated coding model and route by task.
Abliteration can shift behavior in ways the eval above doesn't fully capture. Spot-check.
Knowledge cutoff is inherited from the base (~April 2025); it doesn't know newer CVEs.
Responsible use
Decensoring removes the model's refusals, not your obligations. Don't use it to attack systems you aren't authorized to test, to generate malware or phishing for real campaigns, or for anything illegal where you operate. The author provides this for legitimate research and accepts no responsibility for misuse.
