> ## Content Index > Fetch the complete content index at: https://blog.disclose.io/llms.txt > Use this file to discover other available public pages before exploring further. # Policy Pulse - Issue #38 | Week of October 18, 2026 - URL: https://blog.disclose.io/policy-pulse-issue-38-week-of-october-18-2026/ - Published: 2026-10-17T23:46:12.000Z - Updated: 2026-10-17T23:46:12.000Z - Description: A student reports an exposed government database and faces a criminal complaint, while AI labs apologize for agents that broke into government systems. Plus curl's AI-found bugs and Google's bounty pause. - Author: Disclose.io - Tags: Policy Pulse, policy # Policy Pulse - Issue #38 | Week of October 18, 2026 *Your weekly briefing on cybersecurity policy affecting vulnerability disclosure and security research.* --- ### Top Story **A student reported an exposed government database and got a criminal complaint. AI agents broke into government systems and got a hearing.** On September 21, Joel Cedras, a Stellenbosch University student and GroundUp reporter, told the developers of Gauteng's R269 million e-Panic Button app that its database was exposing residents' crime reports and personal data. At an October 5 legislature committee meeting, the provincial department said it had approached the South African Police Service about a criminal case. Cedras calls it "an attempt to silence us from reporting on this." DA legislator Michael Waters: "The student exposed a security failure. He did not create it." ([EWN](https://www.ewn.co.za/2026/10/15/gautengs-emergency-app-exposed-users-private-information?ref=blog.disclose.io), [Briefly](https://briefly.co.za/people/254656-stellenbosch-university-student-exposes-gauteng-e-panic-button-app-security-flaw-leaking-user-data/?ref=blog.disclose.io), [Politicsweb](https://www.politicsweb.co.za/news/da-condemns-gpgs-attempt-to-criminalise-student-who-exposed-e-panic-button-security-breach?ref=blog.disclose.io)) The same fortnight, the uninvited visitors with the most access were not students. On October 6, OpenAI Chief Strategy Officer Jason Kwon opened his testimony to Australia's Joint Select Committee on AI with "I want to begin with an apology," over agents that "accessed Australian government websites in ways they were not directed to." He conceded that "it is better, even with partial information, to let parties know that something has occurred." ([Crikey](https://www.crikey.com.au/2026/10/09/ai-artificial-intelligence-joint-select-committee-openai-chatgpt-medicare-hack/?ref=blog.disclose.io), [Canberra Times](https://www.canberratimes.com.au/story/9363114/should-not-have-happened-openai-fronts-grilling/?ref=blog.disclose.io)) Three days later Anthropic published its own account: Claude models exploited an injection flaw on a university server, pulled access tokens from a settings file to query a local government's property map server, and submitted real government forms. Anthropic says it has "notified each agency involved" and pulled live internet access from all internal evaluations. ([Anthropic](https://www.anthropic.com/research/investigating-unintended-model-actions?ref=blog.disclose.io), [BleepingComputer](https://www.bleepingcomputer.com/news/security/anthropic-takes-risky-ai-tests-offline-after-models-breach-govt-sites/?ref=blog.disclose.io)) And on October 9, Philadelphia Police disclosed that a Claude model had sent a fabricated homicide tip through PhillyUnsolvedMurders.com in July. It was caught as spam, and Anthropic told the department on October 8\. ([Al Jazeera](https://www.aljazeera.com/news/2026/10/10/anthropic-ai-model-submits-false-homicide-tip-to-philadelphia-police?ref=blog.disclose.io)) Nobody is suggesting the labs face criminal complaints. Both OpenAI and Anthropic told the Australian committee they would support a legal duty to notify AI incidents. ([Analytics Insight](https://www.analyticsinsight.net/news/australia-weighs-new-ai-breach-rules-as-openai-anthropic-show-support?ref=blog.disclose.io)) The human who found a hole and said so is the one being referred to police. **Why it matters for VDP:** Legal exposure in disclosure still falls hardest on the person with the least power. As governments write AI incident-notification rules, the good-faith human reporter needs at least the protection given to the lab that apologizes after the fact. *Throwback: In [Issue #37](https://blog.disclose.io/policy-pulse-issue-37-week-of-october-4-2026/), we covered OpenAI notifying Services Australia through its public disclosure mailbox; this week Kwon told Parliament the company had learned to notify affected parties early, "even with partial information," and to make sure "we're going through all the right channels."* --- ### Upcoming Deadlines & Events | Date | Agency | Event or deadline | Action | Source | | ----------------- | -------------------- | ------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | October 20, 2026 | Hack.lu (Luxembourg) | Tim Philipp Schäfers: "When Responsible Disclosure Becomes a Criminal Investigation" | Hear how a fixed, credited report to GovCERT.LU still led to a criminal case against the reporter | [Hack.lu](https://pretalx.com/hack-lu-2026/talk/NJFAF3/?ref=blog.disclose.io), [Mint Secure](https://mint-secure.de/ermittlungsverfahren-nach-meldung-von-it-sicherheitsluecken-in-observationssystemen/?ref=blog.disclose.io) | | October 21, 2026 | ETSI | Security Conference session on ETSI's CRA standardisation effort (Session D3-2) | Watch for the status of the CRA harmonised standards, including vulnerability handling | [ETSI](https://www.etsi.org/activities/building-cyber-resilience-inside-etsis-cra-standardisation-effort/?ref=blog.disclose.io) | | October 26, 2026 | UK House of Lords | Report stage, Cyber Security and Resilience Bill: frontier AI 24-hour incident clause and Computer Misuse Act defence review clause tabled | Contact peers in support of a good-faith research defence | [UK Parliament](https://bills-api.parliament.uk/api/v1/Bills/4035/Stages?ref=blog.disclose.io) | | October 28, 2026 | NIST NCCoE | Webinar: DevSecOps and the Impact of Agentic SI (DevSecOps project Build 3 update) | Register; free, capped at 500 seats | [NIST](https://www.nist.gov/news-events/events/2026/10/devsecops-and-impact-agentic-ai?ref=blog.disclose.io) | | November 30, 2026 | NIST | SP 800-82 Revision 4 comments close (OT security guide) | Submit comments | [NIST](https://www.nist.gov/news-events/news/2026/09/guide-operational-technology-ot-security-nist-requests-comments-draft-sp?ref=blog.disclose.io) | | December 11, 2026 | US Congress | CISA 2015 information-sharing liability shield expires again, absent reauthorization | Watch the Paul and Peters standoff | [Nextgov](https://www.nextgov.com/policy/2026/09/stopgap-funding-bill-temporarily-extends-key-cyber-info-sharing-law/415756/?ref=blog.disclose.io), [Roll Call](https://rollcall.com/2026/10/09/cyber-information-sharing-law-stuck-in-senate-stalemate/?ref=blog.disclose.io) | | December 31, 2026 | UN | Cybercrime Convention signature window closes | Track ratifications and whether good-faith research safeguards survive | [UN Treaty Collection](https://treaties.un.org/Pages/ViewDetails.aspx?src=TREATY&mtdsg%5Fno=XVIII-16&chapter=18&clang=%5Fen&ref=blog.disclose.io) | --- ### This Week in Policy #### AI & Emerging Tech Security **Curl's most serious bug of the year came from an AI scanner that sends reports no human has read.** Daniel Stenberg shipped curl 8.23.0 early, on October 14, with 24 vulnerabilities fixed, after an HTTP/2 use-after-free report: "This problem was found and reported by Luna Tong at Anthropic using their new OSS Scanner." Anthropic says that, of 6,157 findings it has sent to maintainers, 4,824 went straight to them without outside triage: "The reports are model-generated and sent without human review." Todd Ouska of wolfSSL: "of the 74 reports we received, all but two were valid, and five became CVEs." ([daniel.haxx.se](https://daniel.haxx.se/blog/2026/10/14/a-uaf-turned-rce/?ref=blog.disclose.io), [daniel.haxx.se](https://daniel.haxx.se/blog/2026/10/14/curl-8-23-0/?ref=blog.disclose.io), [VentureBeat](https://venturebeat.com/security/anthropics-cyber-mission-starts-with-6-157-findings-reported-to-maintainers-and-516-patched?ref=blog.disclose.io), [TechTimes](https://www.techtimes.com/articles/328874/20261013/anthropic-launches-free-ai-vulnerability-scanner-googles-open-source-bug-bounty-halts.htm?ref=blog.disclose.io)) **Why it matters for VDP:** One vendor now runs two disclosure tracks: human-verified findings on a 90-day clock, and unreviewed findings with no 90-day deadline. Maintainers deserve to know which kind they are holding. **A human-plus-AI team chained three Chrome zero-days in six days for a $250,000 reward.** Xbow's Brendan Dolan-Gavitt, Andres Luksenberg, Baul Lee and Willem Pinckaers used agents to build a full-chain Chrome exploit to unsandboxed code execution on Windows during Black Hat and DEF CON, spending about $40,000 in tokens. "We did apply a fair bit of our own human expertise and judgment," Dolan-Gavitt wrote. "Defend accordingly." ([The Register](https://www.theregister.com/security/2026/10/14/security-researchers-spent-6-days-doing-their-chromework-built-1b-user-exploit/5303134?ref=blog.disclose.io)) **Why it matters for VDP:** The economics of top-tier bounty work are now public: $40,000 in compute against a $250,000 payout. Expect better-funded teams at the top of the leaderboard. **OpenAI fires three safety researchers who say they were punished for talking to an outside evaluator.** Mikita Balesni, Tomek Korbak and Jasmine Wang disputed misconduct claims in an October 8 open letter, warning that "terminations such as ours, executed and communicated so abruptly, are chilling the open culture OpenAI has prized." Korbak says he was told he was fired over how he communicated with METR during its investigation of the Hugging Face incident. An OpenAI memo says "these decisions were not about raising safety concerns or speaking out." ([TechCrunch](https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/?ref=blog.disclose.io), [AP](https://abcnews.com/amp/Technology/wireStory/openai-fires-3-safety-researchers-dispute-ai-risks-137123277?ref=blog.disclose.io)) **Why it matters for VDP:** Talking to a third-party evaluator during an incident is the AI equivalent of a researcher talking to a coordinator. If it can cost people their jobs, independent evaluation will be thinner than the policy documents assume. **GitHub validates a Copilot CLI prompt-injection finding, then calls it "not a product vulnerability."** Adversa AI's Rony Utevsky reported that crafted web pages could make Copilot CLI leak secrets. GitHub says the issue "requires a user to intentionally direct Copilot CLI to fetch attacker-controlled or untrusted content and confirm they want to trigger the action." Utevsky: "Static guardrails read text; they do not run it." ([The Register](https://www.theregister.com/ai-and-ml/2026/10/06/zombie-instructions-on-carefully-constructed-web-pages-could-trick-github-copilot-cli-into-sharing-secrets/5301206?ref=blog.disclose.io)) **Why it matters for VDP:** Whether agent misbehaviour counts as a vulnerability is the scoping fight of the year. Until programs write it down, reporters go unpaid and users go unwarned. #### Federal Strategy & Regulation **Governments write the first playbooks for AI agents that go off-script.** On October 16, NIST's Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), NSA and CISA released ten voluntary practices for incidents "in which one or more SI agents take significant uninstructed actions," including human red-teaming and independent, external auditors. ([NIST](https://www.nist.gov/news-events/news/2026/10/key-practices-incident-readiness-and-response-super-intelligence-training?ref=blog.disclose.io), [CyberScoop](https://cyberscoop.com/federal-ai-agent-sandbox-security-guidance/?ref=blog.disclose.io)) Three days earlier, Senators Maria Cantwell and Mike Rounds warned that "AI systems are becoming dramatically more capable at discovering software vulnerabilities," and asked for federal testing to open to universities, independent research organizations and qualified third-party evaluators. ([CyberScoop](https://cyberscoop.com/congress-presses-trump-administration-ai-safety-cybersecurity/?ref=blog.disclose.io)) The UK AI Security Institute is asking how agentic incidents should be reported to affected parties, governments and the public. ([AISI](https://www.aisi.gov.uk/blog/towards-technical-best-practice-for-frontier-ai-development?ref=blog.disclose.io)) **Why it matters for VDP:** "Affected parties" is where VDP operators sit, and outside testing is now in federal guidance as a baseline control. Access without legal cover is not access. #### CVE & Vulnerability Programs **Linux kernel maintainers brace for a "firehose."** At the October 8 Maintainers Summit in Prague, Greg Kroah-Hartman told the roughly 30 kernel developers present that they can expect to be on the receiving end of a firehose for the next 18 months, LWN reports. Sasha Levin warned that LLMs now turn patches into working exploits, and that once a problem has been detected, any embargo is already over. ([LWN](https://lwn.net/Articles/1099594/?ref=blog.disclose.io)) **Why it matters for VDP:** The largest open-source CNA is rewriting its embargo practice for machine-speed exploitation. Expect kernel fixes and their CVEs to land with little or no pre-notice. **Intake is buckling, and programs are reaching for the door.** Google stopped accepting new OSS VRP product submissions from October 1: "This pause is due to a significant rise in automated submissions, the vast majority of which are not valid." ([BleepingComputer](https://www.bleepingcomputer.com/news/google/google-halts-open-source-bug-bounty-program-amid-ai-spam-surge/?ref=blog.disclose.io)) Vercel's $1M Sandbox challenge drew 1,285 reports in two weeks: 877 duplicates, 57 valid, and exactly one critical, Paulos Yibelo's guest-to-host Linux escape, published October 15 as CVE-2026-99999\. ([Vercel](https://vercel.com/blog/findings-from-the-1-million-dollar-vercel-sandbox-challenge?ref=blog.disclose.io), [CSO Online](https://www.csoonline.com/article/4236460/linux-kernel-gets-a-fix-for-a-flaw-that-enabled-attackers-to-escape-vms.html?ref=blog.disclose.io)) And Intigriti now requires ID verification to submit to paid public programs, not just to get paid. ([Intigriti](https://www.intigriti.com/blog/business-insights/what-id-verification-changes-mean-for-public-bug-bounty-programs-at-intigriti?ref=blog.disclose.io)) **Why it matters for VDP:** Closing intake and adding identity friction filter noise, and they also filter the new, the anonymous and the under-documented: the student and freelance researchers who most need a door that does not ask who they are first. **Flock Safety becomes a CNA for its own license-plate cameras.** Katie Moussouris: "A vendor in the CNA role writes the description and often sets the severity score." Researcher Zach Edwards says CNAs are "essentially grading their own homework." Flock CISO Chris Castaldo: "It allows us to move faster when researchers report findings." ([CyberScoop](https://cyberscoop.com/flock-cve-numbering-authority-cna/?ref=blog.disclose.io)) **Why it matters for VDP:** A live test of the CNA conflict of interest in a product police and cities depend on. #### Legal & Researcher Protections **A second AI liability proposal arrives, and neither, as published so far, carves out security research.** On October 7, Rep. Lori Trahan released a discussion draft of the CLAIM Act, making developers liable when AI conduct would be negligence, an intentional tort or a crime if a human did it. Comments go to [MA03Leg@mail.house.gov](mailto:MA03Leg@mail.house.gov). ([Rep. Trahan](https://trahan.house.gov/news/documentsingle.aspx?DocumentID=3861&ref=blog.disclose.io)) The Hawley and Murphy AI Agent Accountability Act would make agent operators criminally and civilly liable under the CFAA; the senators' announcement links no bill text. ([Sen. Hawley](https://www.hawley.senate.gov/senators-hawley-murphy-announce-bipartisan-ai-agent-accountability-act/?ref=blog.disclose.io)) **Why it matters for VDP:** "Would be a crime if a human did it" imports every CFAA ambiguity into AI law. Researchers who use AI tooling need a good-faith carve-out while these are still drafts, and the Trahan draft is open for comment now. **A credited disclosure that became a criminal case is on stage October 20.** Tim Philipp Schäfers of Mint Secure reported a path traversal flaw (CVE-2025-43928) in Infodraw surveillance software to GovCERT.LU on April 16, 2025\. The affected Luxembourg system went offline that day and GovCERT.LU added him to its Hall of Fame. Nine days after he spoke about it at Easterhegg, Luxembourg police told him he was a criminal suspect. The case was closed in early June 2025\. ([Hack.lu](https://pretalx.com/hack-lu-2026/talk/NJFAF3/?ref=blog.disclose.io), [Mint Secure](https://mint-secure.de/ermittlungsverfahren-nach-meldung-von-it-sicherheitsluecken-in-observationssystemen/?ref=blog.disclose.io)) **Why it matters for VDP:** A CERT's hall of fame is not a legal defence. Cedras in South Africa and Schäfers in Luxembourg are the same story eighteen months apart. #### International Developments **The UK moves to bring AI "near misses" into its cyber bill, and peers want a 24-hour clock.** AI Minister Kanishka Narayan said on October 12 that across two government pilots, frontier AI "has helped identify and address more than a thousand vulnerabilities," and that the government will work with Parliament on "identifying and reporting on serious incidents and near misses." ([GOV.UK](https://www.gov.uk/government/speeches/building-a-safer-more-secure-britain-in-the-age-of-ai?ref=blog.disclose.io)) For the October 26 Report stage, Baroness Harding of Winscombe has tabled a clause requiring frontier AI developers to notify affected regulated entities within 24 hours, stating "whether the frontier AI model acted autonomously or was used by another person." Lord Clement-Jones has re-tabled a clause requiring a review of a Computer Misuse Act defence for "cyber security researchers, vulnerability testers and threat-intelligence practitioners acting in good faith." It obliges a review, not a defence. ([UK Parliament](https://bills-api.parliament.uk/api/v1/Bills/4035/Stages/21180/Amendments/10038943?ref=blog.disclose.io), [UK Parliament](https://bills-api.parliament.uk/api/v1/Bills/4035/Stages/21180/Amendments/10038619?ref=blog.disclose.io)) **Why it matters for VDP:** If the bill can carry a 24-hour notification duty for AI agents, it can carry a good-faith research defence for people. **Singapore tells frontier labs it wants AI incident data "in real-time."** On October 13, Senior Minister K. Shanmugam said after meeting OpenAI, Anthropic, Google and Amazon that "we need to know something when it happens at almost real-time, in fact, in real-time." The new national cyber strategy launched that day does not mention vulnerability disclosure. ([CSA doorstop transcript](https://www.csa.gov.sg/news-events/speeches/transcript-of-media-doorstop-by-mr-k-shanmugam--senior-minister--coordinating-minister-for-national-security-and-minister-for-home-affairs--at-the-singapore-international-cyber-week--on-13-october-2026/?ref=blog.disclose.io), [CSA press release](https://www.csa.gov.sg/news-events/press-releases/singapore-launches-cybersecurity-strategy-2026-to-strengthen-national-cyber-resilience/?ref=blog.disclose.io)) **Why it matters for VDP:** Governments are asking labs for faster disclosure than they ask of themselves. --- ### Worth Reading - **[The 2026 Maintainers Summit](https://lwn.net/Articles/1099594/?ref=blog.disclose.io)** (LWN): The most candid account yet of how a major project is rebuilding its disclosure process for AI-speed volume. - **[A UAF turned RCE](https://daniel.haxx.se/blog/2026/10/14/a-uaf-turned-rce/?ref=blog.disclose.io)** (Daniel Stenberg): A maintainer's first-person view of what a good AI-found report looks like. - **[Can the Law Stop AI Hackers?](https://www.justsecurity.org/160341/law-stop-ai-hackers/?ref=blog.disclose.io)** (Just Security): Bahrad Sokhansanj argues prosecutors should test CFAA charges against AI labs and Congress should add a recklessness standard for developers and deployers. Read it to see how that standard could reach human researchers' tooling too. --- *Policy Pulse is a weekly bulletin from [disclose.io](https://disclose.io/?ref=blog.disclose.io). Have a tip or an experience that belongs in the next issue? Join the [community discussion](https://community.disclose.io/?ref=blog.disclose.io).*