A quieter news day on launches, but a loud one on the costs of AI in production: capacity limits, security noise, and a model that found a real bug before attackers did.
- Google locks free Gemini users to 3.5 Flash-Lite from Oct 9
- Google pauses part of its bug-bounty program after an AI-slop flood
- Anthropic's Mythos found a Rejetto HFS flaw — exploited within a day
- Bessent calls AI doom warnings "alarmism without solutions"
- Meta's Muse Spark co-authored six math papers
- OpenAI deploys Jalapeño inference ASICs alongside AMD EPYC "Turin"
- Tools worth your time
Google locks free Gemini users to 3.5 Flash-Lite from Oct 9
Starting 9 October, free Gemini users lose access to Flash and Pro and are limited to 3.5 Flash-Lite, with model access adjusted on paid tiers too, according to 9to5Google.
The change reads less like a product decision and more like a capacity one: Flash-Lite is the cheapest model to serve, and routing free traffic there protects margins on the frontier tiers. For anyone who built a habit — or a workflow — on free Flash or Pro, it's a reminder that access on an unpaid tier is a courtesy, not a contract. Worth watching whether rate limits or context windows shift alongside the model downgrade.
Read the full story → https://9to5google.com/2026/10/03/gemini-model-limits-oct-26/?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
Google pauses part of its bug-bounty program after an AI-slop flood
Google suspended product-vulnerability submissions to its OSS VRP bug-bounty program after a wave of AI-generated, hallucinated reports overwhelmed maintainers, Tom's Hardware reports.
This is the predictable downside of lowering the cost of plausible-sounding text: when generating a vulnerability report is nearly free, the signal-to-noise ratio collapses for the humans who have to triage them. It's also a structural problem, not a one-off — any open submission channel that pays out is now a target for automated low-quality volume. Expect more programs to add friction, reputation gating, or paid-stake mechanics in response.
Read the full story → https://www.tomshardware.com/tech-industry/artificial-intelligence/google-suspends-part-of-the-oss-vrp-bug-bounty-program-due-to-an-influx-of-invalid-ai-submissions-product-vulnerability-submissions-ended-october-1?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
Anthropic's Mythos found a Rejetto HFS flaw — exploited within a day
Anthropic's Mythos model identified a critical remote-code-execution flaw in the Rejetto HTTP File Server, and the vulnerability was exploited in the wild within 24 hours of disclosure, according to The Register.
Set against the previous story, the pairing is instructive: AI both floods defenders with false reports and finds real, serious bugs. The sharp detail here is the timeline — disclosure to in-the-wild exploitation in a day narrows the patch window to almost nothing. If model-assisted discovery accelerates on both sides, the advantage goes to whoever can ship and deploy fixes fastest, not whoever finds the flaw.
Read the full story → https://www.theregister.com/security/2026/10/03/anthropics-super-bug-hunting-model-mythos-is-hardcore-good-at-math-as-latest-vuln-under-attack-shows/5300933?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
Bessent calls AI doom warnings "alarmism without solutions"
US Treasury Secretary Scott Bessent dismissed AI-risk warnings from figures such as Dario Amodei, Sam Altman and Elon Musk as "alarmism without solutions," reaffirming his opposition to new federal AI rules, The Next Web reports.
The framing matters because it signals where federal policy is likely to sit: skeptical of safety-first regulation, and inclined to leave rulemaking light. For companies, that reduces near-term compliance uncertainty at the US federal level, while leaving the regulatory action to states and to the EU. Whether "no new rules" holds through an administration is a separate question — but the current direction of travel in Washington is clear.
Read the full story → https://thenextweb.com/news/bessent-says-ai-alarmism-without-solutions-is-not-leadership?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
Meta's Muse Spark co-authored six math papers
Meta says its Muse Spark model co-authored six mathematics papers with human researchers, five of which resolve previously open problems, per Meta AI Research.
Co-authorship with named mathematicians is a stronger claim than a benchmark score: it implies the model contributed to results that peers judged publishable. The caveat is that this is Meta describing Meta's own system, so the external validation to watch for is how the broader mathematics community receives the proofs. If the results hold up to scrutiny, it's a concrete data point on AI as a research collaborator rather than a demo.
Read the full story → https://research.meta.ai/blog/solving-open-research-problems-together?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
OpenAI deploys Jalapeño inference ASICs alongside AMD EPYC "Turin"
OpenAI's hardware VP described deploying the company's in-house Jalapeño inference ASICs with AMD EPYC "Turin" CPUs as hosts, and called Nvidia's standalone Vera CPU "a little bit behind on that maturity level," Tom's Hardware reports.
Custom inference silicon is how the largest operators try to escape Nvidia's pricing and supply constraints, and pairing it with AMD host CPUs is a pointed vote of confidence in a second supplier. The comment on Vera is one vendor's opinion, delivered by a party with an interest in its own stack — treat it as positioning, not a benchmark. The signal worth keeping is the direction: more of the inference stack moving in-house at the frontier labs.
Read the full story → https://www.tomshardware.com/pc-components/cpus/openais-jalapeno-asics-are-deployed-alongside-amd-epyc-turin-cpus-as-hosts-hardware-vp-says-nvidias-vera-standalone-is-a-little-bit-behind-on-that-maturity-level?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
Tools worth your time
- Cactus Whistle — a 16.9MB speech-to-text model that Cactus says outperforms the much larger Whisper Base on several benchmarks while running on CPU. For developers who want on-device transcription without a GPU or a cloud round-trip. https://cactuscompute.com/?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
- Meta Muse Spark — the research model behind today's math co-authorship story, aimed at working alongside researchers on open problems rather than producing finished answers. For mathematicians and research teams testing where model assistance actually holds up. https://research.meta.ai/?utm_source=mother&utm_medium=mother&utm_content=4f4043e7-5b9e-489b-8a71-3b7684fc1499
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
