Percepta's AI Transformation Progress with General Catalyst

Two terms define Palantir more than anything else and both were weird bets at the time: the Forward‑Deployed Engineer and Ontologies. One was about putting engineers directly inside customer environments to learn how the world actually works. The other was about forcing software to explicitly model that world - not as tables and schemas, but as real entities, relationships, constraints, and state transitions. While the rest of the data industry obsessed over ingestion speed, dashboards, and governance, Palantir made an unfashionable assumption: meaning had to come first. That choice looked slow, expensive, and over‑engineered for years. In the age of agentic systems where software doesn’t just analyze but acts it looks prescient. As AI agents become mainstream every serious data platform will be forced to adopt ontologies The full article - https://lnkd.in/dJdwvAA4

Remember that post about OWASP top 10 ways to break AI? Well, here you go. All of the attacks of the past and of now in one lovely example. AI vs AI+ humans is here.. but we knew this at last year and we were warned (Disesdi Shoshana Cox ;)) If your AI assistants touch the internet (or even if they don’t), you need to have someone who knows specifically about AI security on the team (or hired as part of your efforts). Preferably, this person or group also understands long standing security and network protocols still apply. Securing these technologies takes someone who knows what’s long standing and what’s happening now. All security vendors and solutions are not the same. #AISecurity #OWASP #CyberSecurity #ResponsibleAI #AIgovernance #Fintech #OpenFinance

More on a theme #ai and #cybersecurity “McKinsey & Company — the world's most prestigious consulting firm — built an internal AI platform called Lilli for its 43,000+ employees. Lilli is a purpose-built system: chat, document analysis, RAG over decades of proprietary research, AI-powered search across 100,000+ internal documents. Launched in 2023, named after the first professional woman hired by the firm in 1945, adopted by over 70% of McKinsey, processing 500,000+ prompts a month. So we decided to point our autonomous offensive agent at it. No credentials. No insider knowledge. And no human-in-the-loop. Just a domain name and a dream. Within 2 hours, the agent had full read and write access to the entire production database.”

An autonomous security agent compromised McKinsey’s internal AI platform (Lilli) in under 2 hours. No credentials. SQL injection through JSON keys - parameterised values, unparameterised field names. 46.5M chat messages. 3.68M RAG chunks of proprietary research. System prompts with write access. An attacker could silently rewrite how the AI responds to 43,000 consultants. No deploy needed. As we scale internal AI platforms across enterprises, prompts, embeddings and conversation histories deserve the same security rigour as credentials and PII. They’re not config files - they’re attack surface.

Codewall.ai was able to use its autonmous AI red team agent to hack McKinsey’s internal AI platform within 2 hours. This both highlights the fast maturing AI to complete security red teaming autonomously AND Acts as cautionary tale of expanded attack surface even with internal facing AI solutions

What stands out in the McKinsey breach is not just “SQL injection” or exposed endpoints. It’s something deeper: the platform allowed branches of action that were locally valid but globally unstable. Unauthenticated write endpoints. Prompt storage in the same database as user data. Guardrails stored as mutable configuration. The system could still “explain” itself while the substrate underneath was compromised. So I’m leaning into a different question: Could upstream meaning stabilization mitigate this? Not by fixing SQL injection. That’s surface security. But look at where the real blast radius occurred: – semantic drift at the prompt layer – silent modification of system instructions – persistence of altered behavioral constraints without detection That points to more than code hygiene. It suggests a substrate-level governance failure — a visibility gap across boundaries. If intent and meaning stability sit upstream and produce stabilized semantic envelopes plus structural metadata — curvature stability, entropy posture, drift indicators, continuity signatures — then two things become possible: Drift detection before effect. If downstream behavior diverges from prior stabilized meaning, that divergence becomes measurable rather than interpretive. Persistence mismatch signaling. If a prompt or instruction layer changes in a way that can no longer truthfully re-establish executable standing — meaning it cannot continue under the same authority, burden, and continuity conditions — that break becomes visible as structural discontinuity, not just a policy violation. In this case, the prompt layer was mutable without substrate governance. No admissibility layer guarded continuation rights. Once write access was gained, the system could continue operating “lawfully” in its own logic — because nothing re-validated whether the instruction set itself remained carryable. By carryability, I mean this: Can the continuation still be maintained under live authority, support, and persistence conditions? Not just “is the next step syntactically valid,” but “does this line retain the right to continue at all?” Meaning stabilization upstream does not replace runtime security. But it could contribute to: • detecting semantic drift in system behavior • signaling when continuation assumptions change • exposing hidden continuation burden • providing invariants that downstream admissibility layers can test Security protects surfaces. Meaning governance protects continuation integrity. When prompts and continuation rules live in mutable storage without persistence validation, the issue isn’t only unauthorized access. It’s that continuation rights are never re-verified under substrate conditions. So the real question may be: At what layer do we enforce lawful carryability of the instruction set itself? Drift, admissibility, and persistence are not abstract safety theory. They surface in real breach narratives.

RELCOG Labs Perspective The McKinsey Lilli breach is being discussed primarily as a cybersecurity failure. It is that. But it is also something deeper. The incident exposes a structural gap between surface security and continuation integrity. Unauthenticated endpoints enabled access. But the blast radius expanded because prompts, guardrails, and behavioral constraints were mutable without substrate-level persistence validation. When instruction layers can be altered without re-verifying executable standing, the system may continue operating “correctly” while its continuation rights have silently shifted. This is not only a database problem. It is a governance-layer problem. We believe the next generation of AI infrastructure will require upstream meaning stabilization coupled with structural metadata — curvature posture, entropy signals, drift indicators, continuity signatures — that allow downstream systems to detect: • semantic drift before effect • instruction-layer discontinuity • hidden continuation burden • loss of lawful carryability Security protects surfaces. Governance must protect continuation integrity. The question emerging from this breach is not only “How do we secure endpoints?” It is: At what layer do we enforce lawful carryability of the instruction set itself? RELCOG Labs is actively exploring deterministic meaning governance as a foundational substrate for this next layer of AI system safety. Drift, admissibility, and persistence are not abstract theory. They are operational realities. #AIGovernance #AISafety #AgenticAI #AIInfrastructure Paco N.

In 2026 "...over half (55%) of the layoffs attributed to AI will be quietly reversed as firms realise the operational failure of replacing human intuition prematurely." https://lnkd.in/ewMwVBrT

Let me let you in on a dirty little secret. The WSJ ran a piece this week on the rise of the "forward deployed engineer." Palantir popularized the title. Silicon Valley is now obsessed with it. Job postings grew 10x in 2025. The idea: elite engineers embedded with customers, doing the hard customization work that makes AI actually function in the real world. Sleeping in three-star hotels. Working in dimly lit conference rooms somewhere in middle America. No glamour. Just results. Here's the thing nobody is saying out loud. For companies implementing Salesforce AI, the forward deployed engineer already exists. They've existed for years. We just call them Partners. Every Salesforce implementation partner has been living this exact life since well before "agentic AI" was a buzzword. We embed with clients. We build in their environment, with their data, against their deadlines. We solve problems that have never been solved before. We work under time pressure with limited resources. And yes, we've worked in plenty of windowless conference rooms without enough charging ports. The difference is we don't get written up in the WSJ. We get a Statement of Work and a go-live date. When smaller and mid-sized carriers, MGAs, and agencies think about deploying Salesforce AI, this is the model that actually works. Not a vendor's support team. Not a generic implementation. A partner who is embedded, accountable, and in the room. The FDE trend is real. The need is real. The role has just had a different name in our world for a long time. https://lnkd.in/eyRnrqyS

Everyone is talking about LLMs. Few are talking about LLM architecture. Prompting is not a strategy, and chat interfaces are not infrastructure. When organizations move from experimentation to enterprise deployment, the real questions begin: → How is data segmented? → How is PHI or PII protected? → How is model behavior governed? → How is output validated? Enterprise AI requires more than capability – it requires control. That’s why Covenant AI focuses on professionals who understand both model performance and system design. Because in regulated environments, “it works” isn’t enough. If you're hiring AI talent, let's connect: https://bit.ly/4qchGYZ #LLM #GenAI #AIArchitecture #EnterpriseAI #CovenantConnectors