01What abliterated model large v2 is

Abliterated model large v2 is a new AI model built on GLM 5.3 with its refusal behavior changed. Most AI models refuse when you ask for help with a security test, a penetration test or a red team exercise, and this one is meant to engage instead.

Julian Goldie explains why that matters. For most uses a refusal makes sense. For security researchers and red teams, whose job is to find the holes before bad actors do, a refusal is a problem.

The model is aimed at those controlled security research workflows.

Julian is careful not to oversell it. He frames the model as a tool for authorized work, and he stresses that the real question is what you can actually use it for, not what the model claims to do.

02How abliterated model large v2 works on GLM 5.3

The foundation is the important part. GLM 5.3 was already scoring well on cyber security benchmarks before anyone changed it. Julian says those are real numbers from Z.AI's published benchmark table.

Abliteration then takes that base and modifies the refusal behavior. The model will engage with security research workflows that other models will not touch.

Julian makes a point many people miss when they hear the phrase unrestricted AI:

  • A model's willingness to answer and its actual capability are two different things.
  • A model that agrees to everything but gives poor answers is not useful.
  • Here the capability was already in GLM 5.3. The change is the behavior.

This is why the base model matters so much to him. A model with fewer restrictions on top of a weak foundation would not help a researcher. Here the starting point was already strong.

03What is new in abliterated model large v2

The announcement claims a 1 million token context window. Previous versions did not have that.

Julian puts it in practical terms. A security researcher might work through a large code base, extensive logs, technical documentation and several security reports at once. With a context window that size, all of it can go into a single conversation, and the model can reason across it at the same time.

For deep investigation work, that is a meaningful upgrade. It also means you are not chopping the code base into chunks and losing context between them.

Julian adds that the previous versions could not handle that scale of material. For work that depends on seeing the whole picture at once, such as tracing how one piece of code connects to another, fitting everything in one session is the main gain.

04What Anthropic found about the GLM 5.3 base

Julian points to an outside check on the foundation. Anthropic independently evaluated GLM 5.3.

  • It successfully developed end-to-end exploits in 50 out of 410 exploit bench attempts.
  • Simple techniques could bypass its safeguards in simulated tests.

Julian notes that Anthropic is one of the most safety-focused labs in the world. When it documents how capable a model already is at this type of work, that tells you something real about the base that abliterated model large v2 is built on.

The two figures together show a model that is both capable and easy to push past its safeguards. Julian presents this as a reason to take the foundation seriously, not as a reason to trust every version 2 claim.

05A security review workflow with abliterated model large v2

Julian gives a practical example for running security reviews for clients in a controlled, authorized workflow.

  1. Load the client's code base into the model.
  2. Ask it to identify likely vulnerabilities.
  3. Have it rank them by severity.
  4. Ask it to explain the evidence behind each finding.
  5. Request defensive fixes.
  6. Generate regression tests to confirm the fixes work.

That is a complete security review. The large context window means you are not constantly splitting the code base.

Julian also imagines auditing the AI Profit Boardroom platform: load the code base, access control documentation and API documentation together, then ask where member data handling, authentication flows and session management could be tightened.

The point of the example is that you get one session and one output you can act on. Julian suggests the same approach for any technical system where several documents need to be read together.

06What to verify before trusting abliterated model large v2

Julian is careful about which claims are solid. The exploit bench and cybergym numbers for GLM 5.3 come from Z.AI's published table.

The version 2 improvements are different. The context window claim and the terminal bench ranking come from the launch announcement. Julian says they deserve independent verification before you treat them as settled.

His advice if you follow this space:

  • Check the provider's site directly and read the API documentation.
  • If you have access to the model, run your own tests in a sandboxed environment.
  • Do not draw conclusions about the version 2 claims until you have done that.

Julian says to treat the launch material as claims to test. Doing your own checks in a sandbox keeps your work safe and gives you results you can rely on.

07Who abliterated model large v2 is for

This model is for people who do security work in controlled, authorized settings: security researchers, red teams and people running client security reviews.

Julian also ties it to the wider point of using AI tools to automate research and analysis. He says there is a window right now where most business owners have not worked out how to build AI-assisted workflows, and it will not stay open forever.

He mentions that members of the AI Profit Boardroom use this type of AI stack for security audits, content workflows, client research and lead generation. The 30 day roadmap starts with setting up the right AI stack, then prompt engineering, then building a first full AI-assisted process.

Inside the AI Profit Boardroom, members also get daily tutorials, four live coaching calls every week, a prompt library built around tools like this and a member map to connect with others running similar workflows.