TL;DR For vulnerability triage, malware analysis or detection rules in Claude Code, work on Opus 5.5 or Sonnet 5.5, the models with the badge. Fable 5.2 isn't covered: the safeguard still fires there and the session ends up on Opus 4.8. And when a message gets flagged, step back with
/rewindbefore you rephrase it.
CVP: Defense Access is the badge that shows up next to some models in Claude's model picker once your organization joins Anthropic's Cyber Verification Program. It means the cyber safeguards get in the way of defensive work less often. Not on every model, though, and your account carries it into Claude Code too, where nothing tells you which one you're on.
What it is
Claude runs real-time safeguards that stop cybersecurity requests, and sometimes they stop legitimate work. CVP is an identity and use-case check: once you're approved, those safeguards are adjusted to your tier. There are three:
| Tier | What it's for | Who can apply |
|---|---|---|
| Defense Access | Security operations, incident response, malware analysis, detection engineering, vulnerability triage and validation | Organizations and individuals |
| Red Team Access | Authorized pen testing, red teaming, exploit development | Organizations only, up to 75 people |
| Specialized Access | Red teaming critical infrastructure | Organizations only, up to 25 people |
On a Team plan there's nothing to assign: once an Owner applies and gets approved, it turns on for every user in the organization. That's why it showed up without you doing anything.
Which models it covers
All three tiers cover Mythos, Opus and Sonnet starting with Opus 5.5 and Sonnet 5.5. Everything else in your picker keeps the standard safeguards:
| Model | Covered by CVP? | When the cyber safeguard fires |
|---|---|---|
| Opus 5.5 | Yes | Claude Code reruns the request on Opus 4.8 |
| Sonnet 5.5 | Yes | It reruns on Sonnet 5 |
| Fable 5.2 | No | It reruns on Opus 4.8 |
| Haiku 5.5 | No | Not among the models with this classifier |
The third row is the one that hurts. Fable is the model you pick for a long, hard audit, and it's exactly the one without the adjustment: security work hits the wall there the way it always did. When that happens, Claude Code reruns the request on another model and the session stays on it. It says so in the transcript (automatically switched from ... after a message was flagged) and doesn't go back until you pick a model with /model.
What you see in Claude Code
There's no badge in the terminal: the v2.1.327 binary has none in /model. You notice the program when something gets flagged. On Opus 5.5, assembled from the v2.1.327 strings, the notice reads:
Opus 5.5's safeguards flagged this session. You may be seeing this for the first time: Opus 5.5 is more capable and has stronger safeguards as a result, which can sometimes flag non-cybersecurity work. [...] Opus 4.8 is answering instead, or you can edit and retry with Opus 5.5.
At the end of the block message, v2.1.327 adds a sentence based on what the API reports about your account:
| Your status | The sentence |
|---|---|
| Defense Access | Your Defense Access in the Cyber Verification Program covers defensive work, such as vulnerability triage and malware analysis. |
| Your organization is approved, but you weren't assigned | Ask your admin for Defense Access in the Cyber Verification Program to reduce these interruptions. |
| Your organization isn't in the program | Ask your admin to apply to the Cyber Verification Program to reduce these interruptions. |
If you see the first one and still got blocked, what you asked for is outside Defense Access, or it's a false positive you can report with /feedback.
How to work with it
1. Check which account Claude Code signs in with
/status tells you. CVP belongs to the approved organization: your Team's claude.ai login has it. A Console ANTHROPIC_API_KEY doesn't, because Claude and the Console are separate organizations and approval doesn't carry over.
2. Pick the covered model
For security work, /model and Opus 5.5. If the session already dropped to Opus 4.8 after a flag, switch back from there.
3. When it's flagged, step back
The safeguard reads the whole conversation, so sending the same message again in the same session usually gets flagged again. Press Esc twice or run /rewind to return to an earlier point and come at it differently.
4. Decide for yourself whether it switches models
In /config, set Switch models when a message is flagged to Ask each time, or put "switchModelsOnFlag": false in your settings. Claude Code then stops at each flag and lets you switch models or edit and retry.
5. If it fires at startup
The first request carries your CLAUDE.md and git status, and in a security repository that alone can trip it. claude --safe-mode starts without CLAUDE.md, skills, MCP servers or hooks: if it doesn't fire that way, the trigger is in there.
What doesn't change
- You don't need CVP for everyday work: reviewing code, patching known issues and triaging alerts work on any model.
- Defense Access doesn't cover pen testing or exploit development: that's Red Team Access, and it still gets blocked.
- It only touches the cyber safeguards. The biology ones and the Usage Policy stay in full, and what's prohibited stays prohibited on every tier: denial of service, mass data destruction, extortion.
- It's for work on your own systems. Security work for customers needs a separate approval, and on a Team plan access covers the whole organization, so you can't fence it off.
- It requires data retention unless you hold a specific exemption, and from December 15, 2026, Defense Access requires phishing-resistant MFA and stops allowing long-lived API keys.
If you don't have it
As an individual you can only apply for Defense Access, from a Pro, Max, Ultra, Team or Enterprise plan, in the Verification Portal: identity verification with an ID and a selfie, a description of your work, and the terms. The target response time is two days. In an organization, an Owner applies once for everyone.
Official docs: Cyber Verification Program · Automatic model fallback · Safety measures flagged a cybersecurity topic
The messages come from the v2.1.327 binary; the covered models and the tiers, from Anthropic's help article. The picker above is redrawn from a screenshot of a Team account.