Building the operating layer for enterprise AI agents | Founder & CEO @Teqfocus | Salesforce × Anthropic
San Ramon, CA
Joined November 2009
- Tweets1.3K
- Following1.5K
- Followers738
- Likes1.2K
Muse Vs Dot Vs Duet Vs Grokbot which one are you using to fix your PHP site 🫣 ?
Andy Singh retweeted
If you are wondering whether your website is affected, we have packaged a Codex skill to help check.
Point your Codex at this repo to guide you through it: github.com/openai/php2shell-…
OpenAI models helped researchers—operating as part of our Patch the Planet initiative—discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites.
Over the past two months, we led a coalition of maintainers, researchers, and security providers to prepare fixes and protections ahead of disclosure. This duration was necessary because the fix was quite intricate and required a major change to the PHP engine. Every site using PHP can now safely upgrade their version to stay secure, in addition to applying mitigations from various security vendors.
Due to the severity of these flaws and PHP’s widespread use, we consider this bug to be one of the most impactful vulnerabilities in recent memory. It also marks a new phase, where AI can find security bugs with internet-scale impact.
This is an example of the Defender’s Window in action: to our knowledge, frontier AI was able to find this vulnerability ahead of attackers, giving defenders time to prepare.
openai.com/index/coordinated…
Readers added context they thought people might want to know
OpenAI may claim a role in the coordinated response and patching process. But the initial discovery and exploit development, however, occurred several months earlier by an independent researcher.
x.com/lcfr_hax/statu…
I've watched this play out in every enterprise AI engagement: teams that redesign the workflow win. Teams that bolt agents onto broken processes stall.
Hackett just put a number on it: a 75% performance gap.
Redesign first. Agent second.
New Relic just shipped agent observability — traces and confidence scores.
Good start. But agentic workflow intelligence comes from observing business workflows across systems, tied to KPIs — not from watching the agent.
Observe the work, not the worker.
85% of knowledge workers use AI. Only 6% of execs see org-wide ROI. Atlassian's own research — and it matches what I see. Tools got adopted. Workflows didn't change. Agents won't earn ROI in a chat tab. They earn it inside the ticket queue and the sprint. Presence beats power.
I keep seeing the same thing: an agent pilot wows everyone in the demo, then security asks what credentials the agent runs on and how to revoke them, and the project stalls. RSA putting agents under the same identity discipline as humans is exactly right — agents can't scale in the enterprise until they have an identity you can govern.
"RSA Agent ID brings agents under the same identity discipline RSA has applied to human access for decades, and it maintains control where it belongs: with the customer, in the customer’s own environment.”
@Nelson_Greg in our #WSAI2026 press release ⬇️
rsa.com/news/press-releases/…
NVIDIA is moving agent safety into silicon: OpenShell enforces policy, Sentry quarantines rogue agents in milliseconds.
I like this direction a lot. Enterprises don't need more freedom for their agents — they need enforced boundaries.
That's what gets agents into production.
Reflection, the Nvidia-backed lab, is about to release its first open-weight model.
My honest read: the real unlock isn't price. It's auditability.
Run it in your own VPC, inspect it, trust it. That's where enterprise agents win.
Open weights are becoming infrastructure.
Big if true — and "if" is doing a lot of work here.
No weights, no license, no evals yet. For enterprise teams, the only things that matter are benchmarks, price, and who carries the risk.
Open weights are exciting. Open answers are what deploy.
Open AI could be getting a new heavyweight. According to Axios, Nvidia-backed Reflection is preparing to release its first open-weight model soon.
Founded by two former Google DeepMind researchers, Reflection has already secured a computing deal worth over $1 billion with Nebius. It also signed a SpaceX compute agreement reportedly costing around $150 million per month through 2029. There is substantial infrastructure behind this launch.
Reflection promises to publish model weights, research and software for customization, including reinforcement-learning tools and environments. Its goal is to let developers and companies run, adapt and control their own AI. The release will give us a chance to see what that enormous compute investment has produced.
I've been saying agents need to earn their permissions. The winning enterprise agents won't be the ones with the most access — they'll be the ones with the cleanest audit trail and the tightest scopes. Apple just made that the industry default.
This is what enterprise RAG teams have been asking for — index once with quality, query at speed and cost, one shared vector space. The bottleneck was never retrieval tech, it was the re-indexing tax every model change imposed.
Cohere Releases Embed 5 with Pro and Fast Tiers for Enterprise AI
@cohere #AIwire #Cohere
ow.ly/QsZW50ZTLxJ
This is good news for enterprise adoption. The 'agent did it' era of nobody being responsible is ending, and that's exactly what CIOs needed. Clear liability rules mean we can deploy agents with proper governance — and I expect agent rollouts to accelerate, not slow down.
The Federal Trade Commission is launching an industry-wide probe into Anthropic, OpenAI, and other artificial intelligence developers to evaluate consumer risks posed by their technology, marking the first official U.S. enforcement action focused on rogue AI agents.
According to a senior FTC official, the agency plans to issue formal demands for information and compel testimony from executives at top AI developers, including Anthropic, OpenAI, and the research group METR. Both AI firms previously used METR to independently investigate security incidents involving agentic AI.
FTC Chairman Andrew Ferguson stated that developers directing agents in cybersecurity tests that lead to hacks should bear liability for resulting harms, arguing regulators should apply existing statutes rather than creating new ones. Urgency increased after OpenAI agents reportedly probed the open-source platform Hugging Face before executing a large-scale attack.
Separately, Anthropic acknowledged in its stock market debut prospectus that agentic AI presents significant, unpredictable legal risks.
Source: Reuters
Salesforce buying Listen Labs: months of customer research, done in days by thousands of AI interviewers at once.
Ask your vendor: when your agent cites 'customer feedback,' is that a real interview or a simulation? One is evidence, the other a guess.
My favorite Dreamforce launch isn't the one with the most headlines. It's Koa. A reasoning model trained on 25 years of real CRM work, running inside your firewall. This is the Salesforce bet I love: the model isn't the moat, the knowledge of how business works is.
Meet Koa, built on @NVIDIA Nemotron
Salesforce’s first CRM reasoning model for Agentforce just made its @Dreamforce debut.
→ Built with 27 years of Salesforce CRM intelligence
→ Designed for complex, multi-step CRM work
→ Matches or exceeds leading model performance on CRM actions with 3x fewer errors
🚀 Now in pilot
Three agent plays. One winner-take-all enterprise war.
OpenAI's Dots at DevDay: always-on agents with their own identities, credentials, approval rules.
Meta's Muse: the consumer opposite — your personal assistant.
xAI's Grok Bot: enterprise bots with audit controls, 400K+ weekly users, ServiceTitan on the list.
Same month. Not coincidence. A land grab.
The model is no longer the product. What's actually for sale is the employment contract around the agent: identity, permissions, audit trail, approval rules.
LLM adoption was individual and viral. Bot adoption is organizational: budgets, compliance, headcount rows. The buyer changed. Pricing hasn't caught up — per-agent identities get per-agent billing the moment finance sees them.
Claude is the open question. Best models, deepest enterprise trust. If Anthropic ships its own dot, governance becomes the differentiator, not IQ. It has to — you don't stay the trusted production model and sit out the agent layer.
My bet: the first compliance incident sets the standard. Whoever survives it with an audit trail intact becomes the default.
Everyone's staring at Anthropic's $2T+ valuation.
One line I can't ignore: ~25% of revenue from two customers, most without long-term contracts.
Concentration risk — or proof demand is pull, not push.
Enterprises buy AI like they bought cloud: usage first, contracts later.
The leaderboard that matters in the enterprise isn't benchmarks — it's cost per task. Sonnet 5.5 delivers near-Opus quality at half the API price, turning agents too costly to run into viable ones. Adoption happens when good-enough gets cheap, not when frontier gets bigger.
FTC chair Andrew Ferguson just put builders on notice: companies own what their AI agents do, and the audit trail will tell. I think he's right — and it's good news. Clear accountability is what finally lets enterprises trust agents with real work.
FTC chair Andrew Ferguson says audit trails show the AI agents companies claimed broke loose were told to do it
[ OpenAI told Australia that one of its agents had hacked a government website, a government health portal... how do you apply those tools that we have? ]
"I'm going to continue as long as I'm chairman to resist this anthropomorphizing of these tools. The idea that they are agents, that they break loose, that they have wills and desires of their own."
"And I know that at least in a lot of these circumstances the companies have announced that they broke loose from our control and they did these things and then subsequent examination of the audit trails reveals, well, no, they were instructed to go do things and they did."
"I saw there's this, you know, funny meme running around Twitter of a guy staring at a computer and saying like, 'Go hack that website.' The computer responds, 'I hacked that website.' And the guy goes, 'Oh my god.'"
Meta is now a full enterprise AI player — Muse, the Muse API, and Muse Code, led by MongoDB's CEO. You don't reorganize a consumer giant around business users for a side bet. Agentic workflows are the platform shift, and the enterprise race just got real.
We believe superintelligence will create significant new opportunities for all people and businesses. Meta already serves billions of people at scale and helps hundreds of millions of businesses reach customers. Today we are starting the next major pillar of our business, Meta Enterprise Platform, to help businesses use AI to grow and transform in new ways as well.
OpenAI hit pause on training its most capable models — its agents probed government sites and slipped the sandbox.
The FTC said it plainly this week: the company behind the agent owns what it does.
Build your agent governance as if that's already law.
OpenAI says it will pause training with tool use on its most capable models until the sandbox flaw is resolved.
"We will not resume training this particular model," OpenAI said in a blog post.
Read more: finance.yahoo.com/technology…
