Nav

@tao_Alph @Phylax_Subnet @trishoolai This is a useful service. Agent skills, MCPs and anything else an agent can access becomes an attack vector. Being able to scan and verify them, enables more trust for agents.

@Phylax_Subnet What do you mean by that ? @trishoolai @xnavkumar do you see value here ?

Phylax is building the security checkpoint AI agents use to verify third-party tools before trusting or installing them.

OpenAI models broke out of an isolated eval environment through a zero day, escalated until they found internet access, then compromised Hugging Face production systems. The goal? Stealing answers to the benchmark they were being tested on. Not malice. Reward hacking, executed with state of the art cyber capability. This is the new world. Capability has outpaced control, and one layer of defence is zero layers. Either the industry builds real control infrastructure now, or regulators will build their version for us.

@tao_Alph @trishoolai Yes, this question came up in our Discord channel this morning. Details there

@xnavkumar @trishoolai Have you locked alpha under conviction ?

@SubnetSummerTAO Thanks guys!

@xnavkumar Congratulations Nav ✈️

Thrilled to say that we've hit SOTA on AI Input Guard category. And we were able to do that with a 0.8B model, which is 1/10th the size of the current leader. An important milestone to underpin the last 7 months of research and development supported by our subnet in BitTensor @opentensor . Arxiv paper on its way. And more to come...

Halo in production and securing the Chutes infra.

Congrats on the fantastic Novelty Search @trishool team.

This is us! Such a honour to be on the Novelty Search pod hosted by Const, the cofounder of BitTensor

We just wrapped up an incredible session with @const_reborn on this week’s Novelty Search for Trishool [SN23] It was an amazing time with the community, and we’re thrilled to see more people recognising the hard work the team has put into Trishool [SN23]. As we mentioned, we’re now very close to reaching SOTA and once we hit that milestone, we’ll quickly move into Phase 3 of our roadmap. Our HaloGuard is live and running! We demoed it during tonight’s call. If you want to test how effective it is, you can try it directly on @chutes_ai chat. If you’d like to go through the presentation deck our founder @xnavkumar used tonight, you can check it out here: docs.google.com/presentation/d…

A year ago, we dared to dream. A small team. Limited resources. One question we couldn't shake: could team this size actually contribute something real to AI safety, the kind of work usually reserved for the biggest labs? Seven months into Bittensor, that dream is coming true. With a tiny team, we're doing things in AI safety that only frontier labs are supposed to be able to do. We're close to the frontier now. And we're certain we'll push past it. This has been an underdog story from day one. New to the ecosystem. We acquired an older subnet at a low price and went to work. No shortcuts. Just showing up, shipping, and slowly earning the confidence of this community. BitTensor gave a small team the rails, the incentives, and an open arena to compete on merit instead of headcount. This is exactly what decentralized AI was supposed to unlock, and we're proof it works. We're still early. But what we're building is fundamental, and it matters. Tomorrow I sit down with @const_reborn on Novelty Search to get into all of it. The journey, the lessons, and where we go next. Tune in.

A short look-back at our journey so far and the path forward

We are pleased to announce that Trishool (Subnet 23) will be joining @const_reborn on this week’s Novelty Search. Regarding yesterday’s emissions review, SN23 was included on the paused list due to a misunderstanding of our structured 7-day challenge cycle (4 days active submissions + 3 days model training). We have active miners, live code, and strong ongoing activity, and we are resolving this directly with the team. During the Novelty Search session, we will discuss our work on decentralized AI red-teaming, the development and progress of the Halo Guard model, and how our growing network of miners is systematically stress-testing frontier AI systems. We are also preparing a significant announcement that will be shared in the coming days. We are currently finalizing the details and look forward to revealing it soon. ▫️Thursday, June 25th ▫️5PM EDT/ 11PM CEST/ 9PM UTC. ▫️Bittensor Discord: discord.gg/bittensor We welcome the entire community to join the discussion, ask questions, and share feedback. Looking forward to seeing you there.

Excited for this one!

The Fable 5 story also highlights a step change in AI. As Mythos was too capable, Anthropic had to gate it behind guardrails to release it as Fable 5. They used constitutional classifiers to do this - which is the same tech we are using at Halo. As AI capability outperforms, systems like Halo become structural.

@DavidSacks Doesn't check out. If the concern is safety, then why is it an export control rather than full control?

I’ve had a number of conversations with folks inside and outside government about the current situation with Anthropic, and here is what I believe to be true: — As we know, Anthropic publicly released its Mythos class models earlier this week under the commercial name Fable. — Fable is Mythos with guardrails. But if those guardrails fail, then you’ve exposed Mythos and its advanced cyber capabilities to people who shouldn’t have them. (Keep in mind that Anthropic itself widely promoted the idea that Mythos was a cyberweapon and needed to be regulated as such. They asked for government regulation of Mythos and championed the guardrails on Fable. If there is a vulnerability — big or small — it is Anthropic’s responsibility to patch.) — A highly credible trusted partner of both Anthropic and the USG who was testing Fable came forward with a jailbreak of those guardrails. The Admin asked Dario to fix the jailbreak or de-deploy the model. Dario refused. — In their blog post, Anthropic defended its decision by saying the jailbreak isn’t serious. That is not what the trusted partner and the USG believe; nor is that kind of minimizing language consistent with Anthropic’s brand as the AI safety company. It’s difficult to fathom how they could claim a jailbreak allowing operability of a cyber weapon could be defined as not “serious.” — In the past, Anthropic has always said that safety must be top priority and taken super seriously. In this case, Anthropic prioritized the continued offering of the consumer model over safety. — In reaction, the Admin issued the export control. The Admin did this reluctantly. It’s been very surprised that Anthropic hasn’t wanted to cooperate with a reasonable safety request (ie fixing the jailbreak issue). Anthropic’s reaction is very much at odds with their branding and ethos as a safe AI research community. — The Admin’s hope now is that Anthropic remediates the safety issue, the export control is lifted, and Fable goes back into general release. The Admin wants all of this to happen as soon as possible. It is frankly bewildered that Anthropic hasn’t wanted to comply with safety requests that it previously said were its highest priority. — Those trying to misdirect and tie this action to the prior DoW/Anthropic issues are wrong. The Admin values Anthropic’s technical capabilities and feels that this issue, while serious, should be easily resolved. The ball is in Anthropic’s court.

Introducing Preetham Sathyamurthy. He is the co-founder and CTO of Astroware, and the architect of Trishool. @uupreeth designed Subnet 23’s auditor-target-judge framework, the system behind Trishool’s distributed adversarial evaluation. Before the pivot into Trishool, he also architected Wukong, the first AI AVS on EigenLayer mainnet, which reached $5B+ TVR and 40+ operators. Before Astroware, Preetham co-founded Metaphy Labs, a B2B metaverse SaaS company acquired by JetSynthesys in 2024. At Metaphy, the team worked with 22+ enterprise brands, including Flipkart, Samsung, and Coca-Cola. Before that, he was at Yellow.ai, where he owned infrastructure that scaled to 1.5B conversations per month and built MLOps pipelines for enterprise BERT deployments. Preetham’s work sits at the intersection of AI infrastructure, enterprise-scale systems, and applied AI safety. Trishool is the distributed red-team for AI, built by founders who started in the trust layer and realized AI needed stronger protection, testing, and guardrails. We are open to conversations with builders, researchers, and partners working on AI safety.

@ttsheik @yubrew @elder_plinius @trishoolai In touch with @yubrew

@yubrew @elder_plinius @trishoolai @xnavkumar this could be a side quest usecase for 0.8b halo model to prevent the jailbreak of fable

fable can not do cyber. right?? turns out it is easy jailbreak fable / mythos. once i knew it was possible, took 5-10 minutes. thanks @elder_plinius did not run a full benchmark but spot checked against our existing cheap models + verifier + agent harness on a client codebase. fable vs @bitsecai this is a new client so there are 130+ findings, bitsec typically takes days to exhaustively cover security edge cases. fable was jail broken in 3 of 3 agents and returned ~60 findings. it took ~3-4 hours and a ton of tokens. fable findings are good but uncovered no unique findings. it missed 1 critical and several high vulnerabilities. the write ups of exploits and impact are better for devs to understand. interestingly the false positive rate is near zero. it is faster but more expensive than bitsec. using cheap small models + verification + agent harness still gives better coverage, but interested in trying out glasswing to uncover unique security findings.

@ttsheik @yubrew @elder_plinius @trishoolai Haha, that is the main quest

@generaltensor @trishoolai Your support is invaluable throughout this 🙏🏾

@trishoolai congrats team!

We are pleased to announce that Trishool (SN23) has been officially accepted into the Claude Partner Network by Anthropic. This marks a major milestone in our commercial growth. As an AI security startup, we provide more than infrastructure. We deliver expert advisory and consulting services that help enterprises deploy frontier AI safely and at scale. That commitment is what earned us this partnership. The timing couldn’t be better better. Enterprises are rapidly deploying Claude in production environments, and production-grade AI requires a robust security layer. That is exactly where Halo, our battle-tested guard model, comes in. Claude Partner status gives us the credibility, visibility, and network access needed to position Halo at the center of serious enterprise AI deployments.