Ethical Stack.

Composer 2.5

Cursor’s current coding model for sustained agent tasks, based on Kimi K2.5 with further Cursor training and reinforcement learning.

Score calculation

(3 + 4 + 6 + 4 + 1 + 8) ÷ 6 = 4.3

6 of six dimensions rated. Unrated dimensions are excluded from the average.

Full scoring method ↗
Setup details

Composer 2.5 model and Cursor inference route. Cursor identifies Kimi K2.5 as its base and describes additional training and reinforcement learning; the model is available through Cursor.

What fits your needs

Composer 2.5 model and Cursor inference route. Cursor identifies Kimi K2.5 as its base and describes additional training and reinforcement learning; the model is available through Cursor.

Long coding tasks

Cursor describes Composer 2.5 for sustained agent work.

See the evidence

Privacy controls

Privacy Mode prevents use of customer data for training under Cursor’s policy.

See the evidence
Tradeoffs to expect

Hosted inference

Composer 2.5 is not documented as downloadable local weights.

See the evidence

BYOK still uses Cursor

Requests still pass through Cursor for prompt assembly.

See the evidence
How to switch

Change model or client

Moving to another model changes inference route and training data.

See the evidence
Practical questions

Check Privacy Mode

Review account and workspace training controls before sending code.

See the evidence

Separate base and post-training

Review inherited Kimi data rights and Cursor’s post-training sources separately.

See the evidence

Energy practices

3/10↑

Cursor publishes Composer performance and prices but no energy or electricity footprint.

See the evidence

Latency and token cost are not equivalent to energy per completed task. No measured electricity, training compute or water report accompanies Composer 2.5.

1 + 1 reporting + 0 electricity + 1 controls = 3/10. This grades documented practices and transparency.

Task energy, emissions, water, compute hardware and clean power remain unknown.

Reporting1/3
Cursor identifies model and deployment but no energy footprint.
Electricity sourcing0/3
No clean electricity sourcing or facility mix is disclosed.
Resource controls1/3
A fast variant and lower token prices are documented, but no physical resource controls or measured savings.

Water disclosure

Water reporting has not been assessed for this choice.

Training data

4/10↑

Cursor names Composer’s base and post-training process, but has not supplied a complete training-data permission record.

See the evidence

Cursor identifies the base model and describes continued training and RL. A separate report says production interactions can supply reward signals. Privacy Mode controls use of customer inputs, not rights in the inherited Kimi training data or all post-training data.

1 + 2 source transparency + 1 permission + 0 creator controls = 4/10, with the permission cap applied.

A complete dataset list, rights audit for Kimi K2.5, full post-training sources and independent permission validation are not published.

Source transparency2/3
Base model and training stages are identified but material datasets are not itemized.
Documented permission1/4
The training process and base are named, but a complete inspectable permission basis is not documented.
Creator controls0/2
Privacy Mode governs new customer inputs, not permission or removal controls for the original model training data.
Data-rights rubric ↗

Your data & privacy

6/10↑

Composer 2.5 requests go through Cursor; Privacy Mode says customer data is not used for training by Cursor or its providers.

See the evidence

Cursor says custom models such as Composer may use inference providers; Privacy Mode provides ZDR and blocks training use. Without it, Cursor may use codebase data, prompts, snippets and editor actions to improve models. BYOK still passes through Cursor’s backend.

Provider exceptions and security review retention still apply under published terms. Composer is not a local inference route in these sources.

Openness

4/10↑

Composer 2.5 is a hosted model available through Cursor; the model itself is not offered as downloadable weights in the reviewed sources.

See the evidence

Cursor identifies the Kimi K2.5 open-source base but adds Cursor training and RL; Composer 2.5 is offered through Cursor products.

The specific license for all Kimi K2.5 components and any future weight release should be checked at source.

Worker protections

1/10↑

Cursor has public hiring pages, including a federal/DoW sales role, but no opened employment, supplier, or worker-outcome policy establishes compensation, protections, or grievance coverage.

See the evidence

Cursor has public hiring pages, including a federal/DoW sales role, but the careers page is recruiting information rather than a labor standard. Product privacy disclosures don’t show pay, contractor conditions, annotation or moderation labor, supplier standards, or employee grievance routes. The federal/DoW role indicates sales intent only and is not a military customer or contract. These findings concern the developer or operator. This product’s use in a particular deployment needs separate evidence.

1 + 0 pay & conditions + 0 worker voice & remedy + 0 evidence & follow-through = 1/10. This grades documented worker safeguards, not a measured rate of mistreatment.

Compensation, hours, employee feedback/grievance channels, contractor/subcontractor coverage, content moderation and model-data workforce conditions, and independent labor results remain undisclosed in the pages reviewed.

Pay & conditions0/3
No opened source reports worker pay, hours, exposure protections, contractor coverage, or worker-level outcomes for Anysphere / Cursor. The reviewed material is limited to Anysphere product privacy and public company materials checked; no worker or contractor disclosure found. Model-provider labor is outside this company assessment.; its silence is an information gap, not evidence of poor conditions.
Worker voice & remedy0/3
The reviewed material for Anysphere / Cursor doesn’t show worker representation, grievance access, bargaining voice, or documented remedies. Public code contributions, customer controls, or creator rights do not demonstrate worker voice across employees and suppliers.
Evidence & follow-through0/3
No reviewed source provides an independent audit or corroborated worker-level outcomes for Anysphere / Cursor’s relevant employees, contractors, and data-work suppliers. The available disclosures are company/project materials rather than independent labor verification.
Workers rubric ↗

Military ties

8/10Not established↑

Cursor publishes a service-wide AUP, but the targeted official review did not establish a named defense customer or deployment.

See the evidence

Cursor’s AUP applies across Cursor tiers and hosted services; it prohibits unlawful or critically harmful acts, terrorism and weapons of mass destruction and allows service suspension. It does not state a general ban on conventional military/intelligence work or harmful surveillance. Cursor’s careers material showing federal/DoW sales hiring is a sales motion, not proof of a contract or deployment. AUP applies to hosted Cursor use, not user-selected third-party or independent models. These findings concern the developer or operator. This product’s use in a particular deployment needs separate evidence.

7 base + 1 safeguard adjustment = 8/10 (3 of 9 safeguard points; every 3 adds 1). No military relationship was established in the cited sources. This ranks above known defense ties without claiming a verified clean record.

Current contracts, end-user deployments, enforcement outcomes and how the AUP applies to local/self-hosted or customer-selected models are not established.

Use restrictions1/4
Cursor prohibits WMD development, terrorism and certain critically harmful conduct, but leaves conventional military/intelligence support and much surveillance unaddressed; this applies to hosted service, not independently routed models.
Enforcement & oversight1/3
The AUP authorizes suspension/termination for violations, but no military-specific review or enforcement outcomes are published.
Relationship disclosure1/2
AUP scope is public and careers reveal public-sector/DoW sales intent, but no named customer, signed contract or deployment is established.

The official AUP and public career material were reviewed; an advertised DoW/public-sector sales role doesn’t confirm a military customer, contract, or operational deployment.

Military rubric ↗

BDS targets

Not listed↑

We did not find Anysphere among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.

See the evidence

Anysphere is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Check the model and host this tool connects to as well as the company providing the tool. These findings concern the developer or operator. This product’s use in a particular deployment needs separate evidence.

This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.

More from Cursor

All sources

Side by side