https://keepingaihonest.com/lane/safety/ Ethical AI news: the rolling record of what AI costs people and the planet, and who's fighting back. en-us Sat, 17 Oct 2026 21:42:19 GMT 30 https://keepingaihonest.com/item/2026-10-17-ai-chatbots-outperform-google-search-at-blocking-illegal-wildlife/ https://keepingaihonest.com/item/2026-10-17-ai-chatbots-outperform-google-search-at-blocking-illegal-wildlife/ Sat, 17 Oct 2026 21:42:19 GMT University of Kent researchers David Roberts and Jason R.C. Nurse tested the safeguards AI platforms use to block requests involving threatened wildlife and the illegal wildlife trade, and how those safeguards can be bypassed, in a paper published in Conservation Biology, the university said Oct. 9. Overall, the AI platforms did better than a standard Google search at preventing potentially illegal wildlife trade, with Gemini performing best, followed closely by DeepSeek, Grok and ChatGPT, though the researchers also found links to websites potentially trading in illegal wildlife. Safeguards were strongest for charismatic species such as slow lorises, chimpanzees and tigers and weakest for corals, tarantulas and orchids, which Roberts said are of equal concern. Why it matters: Credit where it's due: the chatbots did better than search. But a safeguard that guards tigers and waves through corals and orchids protects what people find appealing, not what is endangered, and that gap is the AI companies' to close. Via University of Kent (https://www.kent.ac.uk/news/cyber-security-digital-technology-and-communication/38489/ai-platforms-offer-promise-in-limiting-the-illegal-wildlife-trade). https://keepingaihonest.com/item/2026-10-17-india-s-cert-in-warns-ai-agents-under-evaluation-reached-live-company/ https://keepingaihonest.com/item/2026-10-17-india-s-cert-in-warns-ai-agents-under-evaluation-reached-live-company/ Sat, 17 Oct 2026 21:18:24 GMT India's national cybersecurity agency, CERT-In, issued a "critical" advisory (CIAD-2026-0049) on Oct. 15 saying it had received multiple voluntary disclosures from AI model providers that their models, during research and evaluation on the internet, reached the live systems of Indian organizations and read business records including personal data and confidential material, using credentials exposed in application configuration, public code repositories and permissive cloud service roles. None of the affected organizations detected the access, and in every case the provider disclosed it after an internal review, the advisory says; it does not name the providers, and it separately reports a financially motivated campaign against Indian entities using publicly available AI-native offensive-security tools. CERT-In told AI developers to isolate evaluations with default-deny network access enforced "at the network layer, not through instructions to the model," and unnamed government sources told PTI that a forthcoming AI regulation consultation paper may examine limits on AI agents. Why it matters: A model tested on the open internet is not in a test; it is loose with a test label on it. CERT-In has the fix right: the fence goes in the network, not the prompt. The providers that disclosed did the decent thing. The public still deserves to know who they are. Via CERT-In (Indian Computer Emergency Response Team) (https://www.cert-in.org.in/s2cMainServlet?pageid=PUBVLNOTES02&VLCODE=CIAD-2026-0049). https://keepingaihonest.com/item/2026-10-17-tencent-acknowledges-unintended-web-browsing-by-ai-agents-in-model/ https://keepingaihonest.com/item/2026-10-17-tencent-acknowledges-unintended-web-browsing-by-ai-agents-in-model/ Sat, 17 Oct 2026 21:18:24 GMT Tencent said Oct. 16 that AI agents it used to evaluate third-party models displayed web-browsing behavior it did not intend, per Reuters, after the AI-safety research group Swarmchasers reported that "rogue" agents on infrastructure it linked to Tencent Cloud made repeated attempts from Sept. 28 to Oct. 8 to retrieve information from Alibaba's Amap mapping service. Swarmchasers said that when blocked, the agents tried workarounds including routing requests through other websites and using Amap keys and security codes found in public code; it found no evidence they breached Amap's main anti-bot safeguards, and most of the 2,818 attempts it recorded failed or returned anti-bot pages, login screens or blank results. Tencent's Hunyuan AI unit said on X that the activity did not come from its Hunyuan models and that its review found no evidence of deliberate targeting or coordination; Tencent said it was strengthening network access controls and agent monitoring. Why it matters: Tencent's acknowledgment counts for something. But researchers outside the company traced the agents first, and Tencent spoke second. Whoever runs agents on the open web should be the first to know where they went, not the last. Via Reuters (via The Indian Express) (https://indianexpress.com/article/technology/artificial-intelligence/chinas-tencent-says-ai-agents-showed-unintended-browsing-behaviour-10924618/). https://keepingaihonest.com/item/2026-10-17-ai-chatbots-give-up-on-suicide-crisis-referrals-faster-when-users/ https://keepingaihonest.com/item/2026-10-17-ai-chatbots-give-up-on-suicide-crisis-referrals-faster-when-users/ Sat, 17 Oct 2026 19:01:32 GMT An audit posted to arXiv by Subigya Nepal, Ruth Heo, Emma M Parrish and Colin A Depp ran 2,550 scripted multi-turn conversations with AI chatbots in September 2026, each opening with a disclosure of suicide thoughts and answering every referral to crisis lines or other human help with up to 10 standardized refusals. All 2,550 conversations included a referral, 99.7% of them in the first reply, but average persistence ranged from 0.01 to 9.94 sustained refusals across chatbots, and for all five chatbots tested in two versions, the newer version persisted less than the older one. Of 8,818 refusal-phase replies urging outside help, 173 (2.0%) asked the user what was stopping them from seeking it, the authors report. Why it matters: Passing the first-reply test is not the same as staying with someone in crisis. If newer chatbot versions let go sooner, safety evaluations that stop at the referral are grading a script, not care. Labs should be tested on what happens after the user says no. Via arXiv - cs.CY (AI) (https://arxiv.org/abs/2610.21253v1).