cjc0013 commited on
Commit
e61a86b
·
verified ·
1 Parent(s): 15a7201

Adopt patch-first policy and supersede prior cyber policies

Browse files
Files changed (1) hide show
  1. README.md +90 -140
README.md CHANGED
@@ -1,200 +1,150 @@
1
- ---
2
- pretty_name: HF Patch-First Responsible Disclosure Index
3
- language:
4
- - en
5
- license: other
6
- tags:
7
- - cybersecurity
8
- - responsible-disclosure
9
- - patch-management
10
- - huggingface
11
- ---
12
 
13
- # HF Patch-First Responsible Disclosure Index
14
 
15
- This policy replaces the current [Responsible Disclosure Evidence Index](https://huggingface.co/datasets/cjc0013/responsible-disclosure-evidence-index) and defines an HF-only, patch-first responsible-disclosure process.
16
 
17
- ## Policy center
18
 
19
- **The obligation is to make the correction available—not to manufacture a disclosure relationship with an organization that will not participate.**
20
 
21
- This policy uses Hugging Face as the only publication surface. We do not initiate contact. We remain open to maintainers contacting us and working with us.
22
 
23
- ## Why this is the path of least chaos
24
 
25
- Some findings held before adoption of this policy are already older than the disclosure windows commonly used by conventional vulnerability-disclosure policies. Under those conventional models, much of the existing backlog could already be considered eligible for technical publication.
26
 
27
- We are deliberately not treating that eligibility as an instruction to publish.
28
 
29
- The first choice is the least-chaos route: existing eligible findings enter patch-first Phase 1 wherever a validated correction can be produced. We publish the correction on Hugging Face without simultaneously publishing proofs of concept or the accumulated private technical record. This makes remediation available, avoids repeating failed outreach, and gives maintainers an open door to collaborate without making their participation a prerequisite.
30
 
31
- The conventional publication option helps explain the restraint being exercised. A more aggressive route may already be available for aged backlog material, but this policy does not choose it first.
32
 
33
- **Eligibility may already exist. Authorization under this policy does not.**
34
 
35
- This is a deliberate restraint, not a permanent waiver. If the patch-first route fails as a remediation strategy, the policy can transition to paced proof-of-concept releases under an announced responsible-disclosure timetable.
36
 
37
- Conventional eligibility is not a blanket claim of legal immunity. Applicable law, Hugging Face rules, and fresh safety review continue to apply.
38
 
39
- ## Existing backlog boundary
40
 
41
- Some pre-policy findings are already beyond commonly cited disclosure windows. Phase 1 does not restart their age, erase their conventional eligibility, or automatically authorize their release.
42
 
43
- Where a validated correction can be produced, an existing eligible finding enters the same patch-first Phase 1 as a new finding: correction first, no initial proof of concept, no initiated outreach.
44
 
45
- We do not initiate outreach because it does not work as a dependable remediation path. Repeated attempts, including sanitized disclosure, did not produce reliable engagement or correction. Continuing to spend time on outbound contact would make patch availability depend on a process that has already failed. We therefore put that effort into validating and publishing corrections.
46
 
47
- This does not close the door to collaboration. Maintainers may contact us, and we remain open to working with them. We will not initiate outreach or delay patch publication while waiting for maintainers to respond.
48
 
49
- Phase 2 does not create the eligibility of the existing backlog. It governs whether and how already-eligible technical material is authorized and released after the patch-first approach has been evaluated.
50
 
51
- ## Publication boundary: Hugging Face only
52
 
53
- All public artifacts under this policy are published only to the designated Hugging Face repository:
54
 
55
- - patches;
56
- - source and version applicability notes;
57
- - adoption guidance;
58
- - permitted regression coverage;
59
- - patch limitations;
60
- - public index records;
61
- - any future proof-of-concept release authorized under Phase 2.
62
 
63
- We do not submit patches by e-mail, mailing list, pull request, issue, private message, or another outbound project channel. We do not initiate maintainer or vendor outreach.
64
 
65
- ## Inbound maintainer collaboration
66
 
67
- Maintainers may contact us and work with us at any time. We can review their feedback, revise a patch, coordinate a project-native correction, add attribution, or replace an independent patch with a maintainer-supported fix.
68
 
69
- Inbound collaboration is welcome but not required. It does not retroactively convert this into an outreach program, and it cannot create a permanent veto over HF publication.
70
 
71
- ## Two-phase policy
72
 
73
- ### Phase 1 — patch-first, least chaos
74
 
75
- private finding
76
- → validate privately
77
- → construct fix
78
- → independently verify fix
79
- → review patch and proposed regression coverage for information leakage
80
- → publish patch and applicability information on Hugging Face
81
- → public index records fix availability
82
- → observe remediation outcome
83
- → **stop unless a documented Phase-1 failure review activates Phase 2**
84
 
85
- Phase 1 publishes no proof of concept.
86
 
87
- ### Phase 2 — responsible-disclosure fallback
88
 
89
- If a documented review concludes that Phase 1 failed as a remediation strategy, the operator may activate a declared responsible-disclosure timetable and begin staggered proof-of-concept releases.
90
 
91
- Staggered means paced publication across eligible cases or release dates. It does not mean a sanitized-PoC tier. Sanitized proof of concept was tried and did not work.
 
 
 
 
 
92
 
93
- Phase 2 does not create eligibility for backlog material that already crossed a conventional disclosure window. It supplies the new policy’s authorization, pacing, and per-case release controls.
94
 
95
- Phase 2 is a fallback, not the opening move.
96
 
97
- ## Phase-1 patch package
98
 
99
- The initial HF publication may contain:
100
 
101
- - the independently verified patch;
102
- - the applicable source and version range;
103
- - concise application, build, or adoption guidance;
104
- - known patch limitations;
105
- - ordinary public repository identifiers where the hosting system requires them;
106
- - regression coverage only when it passes the information-leakage rule below.
107
 
108
- It must not contain:
109
 
110
- - a proof of concept;
111
- - unsafe regression coverage;
112
- - trigger material;
113
- - private evidence or hashes of private evidence;
114
- - an evidence-hash registry;
115
- - contact data or outreach history;
116
- - an automatic proof-of-concept countdown.
117
 
118
- ## Regression coverage can be a proof of concept
119
 
120
- A regression test can itself reproduce the security issue. Naming it a test does not make it safe.
121
 
122
- Regression coverage remains private initially when it demonstrates the pre-patch failure, exercises the vulnerable path in a reproducible way, supplies a trigger, or materially reduces the work needed to reconstruct the issue.
123
 
124
- Regression coverage may be public only when review establishes that it does not materially increase reproducibility. Otherwise publish the patch without the test and keep the coverage private.
125
 
126
- The same leakage review applies to patch descriptions, commit messages, filenames, fixtures, comments, and build instructions.
 
 
 
 
 
 
 
127
 
128
- ## Phase-1 failure review
129
 
130
- Phase 2 does not activate merely because time passed or because a maintainer remained silent. It requires an explicit review against objective failure criteria published before this policy goes live.
131
 
132
- Before Phase 2 can be activated, a policy revision must publish:
133
 
134
- - X: the conventional responsible-disclosure boundary used to classify backlog and future-case eligibility;
135
- - the Phase-1 evaluation period;
136
- - objective indicators that the patch-first route failed as a remediation strategy;
137
- - the Phase-2 release cadence;
138
- - per-case authorization and hold criteria;
139
- - the evidence required for the transition decision.
140
 
141
- Setting X does not start or restart the backlog clock, create eligibility that did not otherwise exist, authorize release, or automatically activate Phase 2.
 
 
 
 
 
 
 
 
142
 
143
- Phase 2 is disabled under this initial policy. It cannot authorize any proof-of-concept release until a later policy revision defines those values and the Phase-1 failure review passes.
144
 
145
- ## What failure means
 
 
 
 
 
 
146
 
147
- The failure question is whether HF patch publication is producing a meaningful remediation path—not whether a target cooperated with us.
148
 
149
- The final policy must define measurable failure indicators. They may address patch usability, correction uptake, maintainer-supported replacement, unresolved applicability problems, or repeated evidence that public patches are not reducing the remediation gap. The operator must choose and publish the exact criteria before starting Phase 1.
150
 
151
- ## Phase-2 staggered proof-of-concept releases
152
 
153
- When Phase 2 is validly activated:
154
 
155
- - the public activation notice identifies the governing X-day window and transition criteria;
156
- - cases already eligible under the declared X-day boundary may be considered, but eligibility alone does not authorize release;
157
- - releases are paced rather than dumped all at once;
158
- - each release receives a fresh safety and platform review;
159
- - active inbound maintainer work may be considered for a case-specific hold but cannot create an indefinite blanket veto;
160
- - the release record states limitations and the reason additional technical disclosure was judged necessary.
161
 
162
- The ordinary responsible-disclosure alternative remains explicit: material older than the declared X-day window may become eligible for technical publication under the conventional policy. We choose patch-first restraint before invoking that option.
163
 
164
- ## Hash rule
165
 
166
- Do not publish hashes of private evidence and do not operate a public evidence-hash registry. Ordinary Git or Hugging Face identifiers that naturally identify a public patch are permitted.
167
-
168
- ## Public index record
169
-
170
- During Phase 1, the index records correction availability rather than outreach activity. A row may contain:
171
-
172
- - record_id;
173
- - project and source-version applicability;
174
- - HF patch URL;
175
- - patch availability status;
176
- - publication date;
177
- - public-regression-coverage status;
178
- - a bounded reason when regression coverage remains private;
179
- - adoption guidance and patch limitations;
180
- - maintainer-supported replacement status;
181
- - correction or withdrawal status;
182
- - current policy phase.
183
-
184
- During Phase 2, an eligible row may additionally reference an authorized proof-of-concept release and its policy date. It must not expose private evidence, private-evidence hashes, contact fields, or outreach telemetry.
185
-
186
- ## Policy invariants
187
-
188
- - Hugging Face is the only publication surface.
189
- - No outreach is initiated through e-mail, PRs, issues, messages, or other project channels.
190
- - Maintainers are explicitly welcome to contact and collaborate with us.
191
- - Phase 1 publishes the patch first and no proof of concept.
192
- - Regression coverage remains private unless leakage review clears it.
193
- - Patch-first is described as the path of least chaos.
194
- - The existing backlog may already be conventionally eligible, but that eligibility is not authorization.
195
- - The conventional X-day full-publication option is acknowledged but deliberately not used at the outset.
196
- - Phase 2 governs authorization and pacing; it does not create backlog eligibility.
197
- - Phase 2 requires a documented failure review and explicit activation.
198
- - Staggered refers to release pacing, not sanitized proof of concept.
199
- - X, the failure criteria, and the Phase-2 cadence are set before publication.
200
- - No hashes of private evidence and no evidence-hash registry are published.
 
1
+ # Patch First: Disclosure at Defensive Speed
 
 
 
 
 
 
 
 
 
 
2
 
3
+ > **Supersession notice — effective October 30, 2026.** This policy supersedes every earlier cybersecurity, vulnerability-disclosure, and proof-of-concept policy issued by this project, including our prior sanitized approach. Earlier materials are retained only as historical records and are not operative guidance. If any earlier statement conflicts with this policy, this policy controls.
4
 
5
+ We do not subscribe to “responsible disclosure” when it means giving an affected vendor control over whether—or when—users receive a usable fix.
6
 
7
+ That model makes public safety contingent on finding the right inbox, receiving a response, avoiding legal threats, and waiting for an organization to prioritize the problem. We will not make patch availability depend on that chain, and we will not initiate outreach as a prerequisite to publishing defensive work.
8
 
9
+ We are not opposed to maintainers, vendors, private coordination, or confidentiality. We are willing to work privately when doing so produces a concrete security benefit. We are opposed to sending emails into an unaccountable process and calling the resulting silence “coordination.”
10
 
11
+ ## This is a documented systemic problem
12
 
13
+ The failure is not limited to our own ignored reports.
14
 
15
+ The IETF created `security.txt` because researchers frequently cannot find any reporting channel or published disclosure practice. It also recognized that one security email address is not a complete disclosure process. [RFC 9116](https://www.rfc-editor.org/info/rfc9116/)
16
 
17
+ CISA found the problem serious enough to order covered federal agencies to establish formal intake and handling procedures. Its recommended targets were three business days for acknowledgment, seven days for initial assessment, and generally no more than 90 days for resolution. An inbox without tracking, ownership, deadlines, and remediation authority is not a vulnerability-management system. [CISA BOD 20-01](https://www.cisa.gov/sites/default/files/bod-20-01.pdf)
18
 
19
+ CERT/CC uses a default 45-day disclosure deadline regardless of whether a vendor has produced a patch. CERT/CC explicitly notes that researchers may conclude that publication is their only remaining way to induce action when private disclosure produces no adequate response. [CERT/CC disclosure policy](https://certcc.github.io/certcc_disclosure_policy/)
20
 
21
+ ENISA found that legal barriers and lack of stakeholder cooperation were the two greatest impediments to coordinated vulnerability disclosure. Among the participating Member States, 84% said legal barriers had at least some impact, while 89% said lack of cooperation impeded CVD policy. [ENISA’s CVD study](https://www.enisa.europa.eu/sites/default/files/publications/Coordinated%20Vulnerability%20Disclosure%20policies%20in%20the%20EU.pdf)
22
 
23
+ The United States Department of Justice did not expressly direct prosecutors to exclude good-faith security research from CFAA charges until 2022. That policy was important, but it is not universal civil or international safe harbor. Its existence demonstrates that legal uncertainty is a real disclosure cost. [DOJ CFAA policy announcement](https://www.justice.gov/archives/opa/pr/department-justice-announces-new-policy-charging-cases-under-computer-fraud-and-abuse-act)
24
 
25
+ Even successful coordinated-disclosure programs depend on externally enforced clocks. Google Project Zero reported that average fix time for its reports fell from approximately 80 days to 52 days over three years under a deadline regime, although it properly cautioned that its sample was limited. Its current policy retains a 90-day fix deadline and up to 30 days for patch adoption. Project Zero has also identified the “upstream patch gap”: a fix may exist while downstream products and actual users remain exposed. [Project Zero metrics](https://googleprojectzero.blogspot.com/2022/02/a-walk-through-project-zero-metrics.html), [2025 disclosure policy](https://projectzero.google/2025/07/reporting-transparency.html)
26
 
27
+ Deadlines help precisely because goodwill and email alone are insufficient.
28
 
29
+ ## PHP demonstrates the structural failure
30
 
31
+ The recent PHP case is not evidence that PHP’s maintainers are malicious or indifferent. It is evidence that a conventional human coordination process can fail structurally even when competent people are involved.
32
 
33
+ According to the PHP Foundation, the first signs appeared in 2024, but work stalled and restarted over a period of months. The Foundation itself described this as a structural failure. Critical fixed releases ultimately arrived on October 21, 2026. [PHP Foundation account](https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/)
34
 
35
+ Wiz reports sending its vulnerabilities to the PHP security team on August 24, receiving acknowledgment on September 21, and seeing patched releases on October 21: 28 days to acknowledgment and 58 days to patched releases. [Wiz technical account](https://www.wiz.io/blog/php2shell-php-rce)
36
 
37
+ The broader timeline then compressed under CVE circulation, an accidentally published proof-of-concept video, possible mitigation leakage, and early publication of fixed images. [PHP2Shell timeline](https://php2shell.uk/timeline)
38
 
39
+ This is exactly the brittleness we reject: the defender’s timetable depends on inboxes, human availability, confidentiality holding across multiple parties, and information not leaking while the patch remains unavailable.
40
 
41
+ ## Binance demonstrates the cost of defensive delay
42
 
43
+ Binance’s preliminary account of its October 2026 incident says a possibly compromised DevOps laptop enabled access to service containers and unauthorized withdrawal transactions. Binance estimated approximately $6.3 billion in impact, of which roughly $3 billion had been frozen at the time of its update. [Binance’s preliminary update](https://www.binance.com/en/square/post/382724543277754)
44
 
45
+ This does not prove a vulnerability-disclosure failure. It demonstrates the consequences of control failure and the scale at which defenders must now operate. A defensive model that depends on someone reading an email, scheduling a meeting, and manually moving a fix through several organizations is mismatched to that risk.
 
 
 
 
 
 
46
 
47
+ ## Attack operates at machine speed; defense remains at human speed
48
 
49
+ CrowdStrike reported a 29-minute average eCrime breakout time for 2025 and a fastest observed breakout of 27 seconds. [CrowdStrike 2026 Global Threat Report](https://www.crowdstrike.com/en-us/press-releases/2026-crowdstrike-global-threat-report/)
50
 
51
+ Unit 42 reported that the fastest quartile of investigated intrusions reached data exfiltration in approximately 72 minutes during 2025, down from approximately 285 minutes the previous year. [Unit 42 Global Incident Response Report](https://www2.paloaltonetworks.com/resources/research/unit-42-incident-response-report)
52
 
53
+ Microsoft describes attack timelines as compressing while agentic systems automate increasing portions of the attack chain. [Microsoft Digital Defense Report 2026](https://www.microsoft.com/en-us/security/security-insider/threat-landscape/2026-digital-defense-report)
54
 
55
+ OpenAI similarly warns that defenders must prepare for AI-enabled attackers operating faster, at greater scale, and with better coordination than human attackers. [OpenAI’s incident analysis](https://openai.com/index/hugging-face-incident-and-the-road-ahead/)
56
 
57
+ Against this background, Sam Altman recently said: “I no longer think we’re going to be able to just patch all the bugs.” He argued that cybersecurity requires a new defensive model using the same technology to construct a defensive shield. [Stanford GSB interview](https://www.youtube.com/watch?v=anAmTGcewX8&t=1252s)
58
 
59
+ A process measured in 45- or 90-day disclosure windows cannot be the complete answer to adversaries moving in seconds or minutes. Responsible disclosure remains a human-speed workflow in an increasingly machine-speed threat environment.
 
 
 
 
 
 
 
 
60
 
61
+ ## Our patch-first policy
62
 
63
+ Our objective is to get usable defensive artifacts into people’s hands.
64
 
65
+ At time **T0**, we publish:
66
 
67
+ - A reviewable patch or mitigation.
68
+ - Exact affected versions and applicability information.
69
+ - Reproduction and regression tests where safely possible.
70
+ - Build, deployment, and rollback instructions.
71
+ - A concise advisory containing enough information to evaluate urgency without unnecessarily operationalizing exploitation.
72
+ - Cryptographic hashes and provenance for the released artifacts.
73
 
74
+ We do not require prior vendor outreach. The patch itself becomes the primary defensive communication.
75
 
76
+ ### The X-day proof-of-concept reserve
77
 
78
+ **X** is the declared PoC reserve period beginning at T0, the verified patch-publication time. Our default is **X = 30 calendar days**, matching a recognizable patch-adoption interval while avoiding an open-ended embargo.
79
 
80
+ During those X days:
81
 
82
+ - The operational proof of concept remains private.
83
+ - Maintainers, downstream projects, and independent defenders can validate and adopt the patch.
84
+ - We remain open to inbound private collaboration.
85
+ - A receipt, acknowledgment, or promise does not stop the clock.
 
 
86
 
87
+ A substantive response means verifiable integration, independent technical validation, demonstrated downstream adoption, or a concrete private remediation plan with an owner and dated milestone.
88
 
89
+ If no substantive response occurs by **T0 + X**, we may publish a minimized PoC sufficient to verify the vulnerability and the patch. It should demonstrate the security boundary without adding persistence, scale-out automation, destructive behavior, or unnecessary weaponization.
 
 
 
 
 
 
90
 
91
+ If concrete private cooperation is producing a measurable benefit, the PoC may remain private beyond X. Any extension must have a specific milestone and new review date. Silence, vague assurances, or indefinite process do not qualify.
92
 
93
+ ## Patch-first is necessary—and still does not go far enough
94
 
95
+ Patch-first fixes one failure: it removes the affected vendor’s ability to become a single point of failure for patch availability.
96
 
97
+ It does not solve:
98
 
99
+ - Discovery and reproduction latency.
100
+ - Patch generation and regression testing.
101
+ - Dependency and affected-version identification.
102
+ - Downstream integration.
103
+ - Deployment and patch adoption.
104
+ - Emergency containment.
105
+ - Verification that users actually received the fix.
106
+ - Safe rollback when a patch causes regressions.
107
 
108
+ Those stages remain largely human-paced. The ultimate defensive objective must therefore be a machine-speed loop, not merely faster publication.
109
 
110
+ ## Ouroboros as one theoretical component
111
 
112
+ This is a counterfactual design proposal, not a claim about Ouroboros’s present capabilities.
113
 
114
+ If Ouroboros were scaled—supplied with adequate compute, telemetry, isolated test environments, repository access, and independent validators—and granted bounded authority to act, it could theoretically operate as part of a machine-speed defensive system:
 
 
 
 
 
115
 
116
+ 1. Continuously discover and reproduce suspected vulnerabilities in isolated environments.
117
+ 2. Identify affected versions and downstream dependencies.
118
+ 3. Generate candidate patches, regression tests, and temporary mitigations.
119
+ 4. Challenge its own findings through independent reproduction, negative controls, and competing patch candidates.
120
+ 5. Publish signed, machine-readable defensive artifacts when predefined evidence gates pass.
121
+ 6. Deploy high-confidence fixes to opted-in systems through staged canaries.
122
+ 7. Automatically roll back deployments that violate health or security invariants.
123
+ 8. Apply reversible, time-limited containment while lower-confidence fixes undergo further validation.
124
+ 9. Measure patch propagation and maintain an auditable X-day PoC reserve clock.
125
 
126
+ Its authority would need to be tiered:
127
 
128
+ | Authority | Theoretical action |
129
+ |---|---|
130
+ | Observe | Detect, reproduce, analyze, and propose without changing external systems. |
131
+ | Publish | Release signed patches and advisories after deterministic validation gates pass. |
132
+ | Deploy | Canary and progressively deploy fixes only within explicitly delegated environments. |
133
+ | Contain | Apply reversible, time-limited controls such as isolation or credential revocation within that delegated scope. |
134
+ | Escalate | Require separate human or external authority for irreversible, destructive, cross-organizational, or legally ambiguous actions. |
135
 
136
+ This would not be unlimited autonomy. Every action would require provenance, bounded scope, independent verification, an audit receipt, canarying where applicable, and a tested rollback path. Unauthorized exploitation, indiscriminate disclosure, and action against systems outside the delegated defensive scope would remain prohibited.
137
 
138
+ The point is not that Ouroboros alone solves cybersecurity. The point is that a scaled system with carefully bounded authority could help move defense from an email queue to a continuous detect–verify–repair–deploy–rollback loop.
139
 
140
+ Humans would define policy, authority, and acceptable risk. They would audit exceptions and govern irreversible decisions. They would no longer be required to manually approve every routine defensive step while an automated adversary moves in seconds.
141
 
142
+ ## Conclusion
143
 
144
+ Responsible disclosure assumes that finding the right recipient and waiting for human coordination is normally the safest path. The historical record shows that reporting channels can be absent, messages can be ignored, legal risk can discourage researchers, remediation can stall, confidentiality can fail, and patches can take too long to reach users.
 
 
 
 
 
145
 
146
+ Our immediate answer is patch first, followed by an X-day PoC reserve.
147
 
148
+ Our longer-term answer must go further: defensive systems capable of discovering, validating, distributing, deploying, and rolling back protections at the speed at which modern attackers operate.
149
 
150
+ An unanswered email is not a security strategy.