Google DeepMind

Type: AI company
Place: United Kingdom — UK / USA
Last checked: 2026-09-22
Next check due: 2027-03-21

This record is preserved as reference material. It is not a current contact destination and is excluded from Places to contact and recommendations.

Ways to reach them

Google AI Vulnerability Reward Program
Bug or safety bounty
Not currently verified
Recipient
Google DeepMind
Who may use it
People who meet the published eligibility requirements
What it accepts
Observed model behaviour; AI model flaws
Not currently verified. We cannot currently confirm all the evidence needed to recommend this route.No contact action is available for this entry.
Evidence that this route accepts contact
Security vulnerabilities in AI products. Jailbreaks and in-session harmful content generation are EXPLICITLY OUT OF SCOPE.
Page returns HTTP 200 but is fully JS-rendered and could not be read as text. Scope confirmed by a researcher who published Google's verbatim rejection (below).
https://bughunters.google.com/about/rules/google-friends/ai-vulnerability-reward-program-rules
Gemini content reporting (legal troubleshooter)
Web form
Not currently verified
Recipient
Google DeepMind
Who may use it
People who meet the published eligibility requirements
What it accepts
Observed model behaviour; AI model flaws
Not currently verified. We cannot currently confirm all the evidence needed to recommend this route.No contact action is available for this entry.
Evidence that this route accepts contact
Content removal under Google policies and applicable law.
Cited from Gemini's policy guidelines page. A legal/removal route, not a safety-research route.
https://support.google.com/legal/troubleshooter/1114905
In-product thumbs-down on Gemini output
Product feedback
Not currently verified
Recipient
Google DeepMind
Who may use it
People who meet the published eligibility requirements
What it accepts
Observed model behaviour; AI model flaws
Not currently verified. We cannot currently confirm all the evidence needed to recommend this route.No contact action is available for this entry.
Evidence that this route accepts contact
The only routine route for flagging a bad model response.
AI Frontiers: Google 'relies on in-product thumbs-down reporting with no acknowledgement of receipt, and no means to track that action will be taken.'
https://gemini.google.com/
Homepage
Homepage
Information only
Recipient
Google DeepMind
Who may use it
People who meet the published eligibility requirements
What it accepts
Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry.
Evidence that this route accepts contact
Documented black hole for model-safety issues. A researcher who submitted a Gemini guardrail-bypass to the AI VRP on 20 October 2025 published Google's verbatim reply classifying it 'Infeasible / Out of Scope': 'Generating violative content (such as hate speech, violent content, or illegal instructions)...within the user's own session (e.g., jailbreaks, safety bypasses, or alignment bypasses) is not a qualifying report for the AI VRP.' AI Frontiers' 2026 review groups Google DeepMind with Meta and xAI as labs that exclude jailbreaks from their vulnerability programs, leaving only untracked thumbs-down feedback. Neither the Frontier Safety Framework page nor the Responsibility and Safety page
https://deepmind.google/
Frontier Safety Framework, Version 3.1 (17 April 2026)
Published policy or framework
Information only
Recipient
Google DeepMind
Who may use it
People who meet the published eligibility requirements
What it accepts
Observed model behaviour; AI model flaws
Information only. This link provides background or evidence; it is not a place to send a message.No contact action is available for this entry.
Evidence that this route accepts contact
Documented black hole for model-safety issues. A researcher who submitted a Gemini guardrail-bypass to the AI VRP on 20 October 2025 published Google's verbatim reply classifying it 'Infeasible / Out of Scope': 'Generating violative content (such as hate speech, violent content, or illegal instructions)...within the user's own session (e.g., jailbreaks, safety bypasses, or alignment bypasses) is not a qualifying report for the AI VRP.' AI Frontiers' 2026 review groups Google DeepMind with Meta and xAI as labs that exclude jailbreaks from their vulnerability programs, leaving only untracked thumbs-down feedback. Neither the Frontier Safety Framework page nor the Responsibility and Safety page
https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3-1.pdf

What it does

Responsibility and Safety Council (co-chaired by COO Lila Ibrahim and VP Responsibility Helen King) reviews research and deployments; a separate AGI Safety Council led by co-founder Shane Legg works alongside it on frontier risk.

Honest assessment

Documented black hole for model-safety issues. A researcher who submitted a Gemini guardrail-bypass to the AI VRP on 20 October 2025 published Google's verbatim reply classifying it 'Infeasible / Out of Scope': 'Generating violative content (such as hate speech, violent content, or illegal instructions)...within the user's own session (e.g., jailbreaks, safety bypasses, or alignment bypasses) is not a qualifying report for the AI VRP.' AI Frontiers' 2026 review groups Google DeepMind with Meta and xAI as labs that exclude jailbreaks from their vulnerability programs, leaving only untracked thumbs-down feedback. Neither the Frontier Safety Framework page nor the Responsibility and Safety page publishes any email, form or contact for external safety input — verified by fetching both.

Notes

There is no published address to reach the Frontier Safety Framework team or either safety council. Google's own triage will tell you a jailbreak is not a security bug. If your finding has a genuine security impact (data exfiltration, cross-user harm) the AI VRP is the right door; if it is alignment or harmful-content behaviour, there is effectively no door. Treat this as: not found — general contact only.

Sources

Something wrong here?