Forward-Deployed Means Shipping the Triage Too

Two data points landed this week that say more about production AI delivery than most launch posts. VulnCheck's Patrick Garrity has published his seven-month audit of Anthropic's Project Glasswing, finding that an AI system can discover vulnerabilities far faster than anyone can validate them S1. In the same few days, IEEE Spectrum argued that agentic systems are straining zero-trust security because nobody can fully verify what autonomous agents actually did S2. Read together, they describe the next constraint on AI delivery: the model is no longer the bottleneck — the receiving side is.

The bottleneck moved, then it hopped the fence

The forward-deployed engineering story of 2026 rests on a simple diagnosis. GFT's global CEO Marco Santos wrote in October that OpenAI, AWS, Anthropic and Microsoft are all betting on embedding engineers with customers because "deployment has become the new frontier" S3. The FDE model works because someone accountable stays with the work until it functions in the customer's environment.

Project Glasswing shows what delivery looks like when that embedded loop is missing. By Anthropic's own November 2 ledger update, the effort had accumulated 37,966 vulnerability findings, of which only 217 were marked fixed; in the two months since September, findings grew by roughly 11,813 while fixes grew by 15 S1. Discovery scaled with the model. Validation stayed human-sized, and the gap between the two is now the project's defining number.

Unverified output is a cost transfer, not a deliverable

The same pattern is visible at the individual maintainer level. Anthropic's self-service OSS Scanner sends reports that are "fully model-generated, without human review or triage," shifting validation onto open source teams. Curl's Daniel Stenberg received 45 such reports; of the 34 published so far, 19 were rejected as not vulnerabilities and 6 confirmed as low severity, and he is deliberately pacing publication to soften the impact on his team S1.

That is the analysis worth internalizing: an AI pipeline that emits artifacts at machine pace does not finish its work when generation completes. It transfers a triage obligation to whoever receives the output. Whether you book it or not, that cost lands on your delivery budget — or on your customer's.

Zero trust has to live where agents act

The same lesson runs through the deployment side. IEEE Spectrum reports that applying zero-trust architecture to agentic systems forces an awkward question: the policy enforcement point itself must be trusted, so placing all verification at one choke point may not scale to autonomous agents. Researchers at Queen's University Belfast found orchestrated agents could execute de-authentication attacks in as little as two to three seconds, and point to AI bills of materials and distributed-ledger traceability as ways to keep verifiable evidence of agent actions S2. Verification has to sit where the agent acts, not wherever it is convenient to audit later.

What production teams should take from this

Three practices follow. First, budget the receiving side: any AI-generated artifact stream — findings, code, reports — needs a named triage owner and a rate the receiver can actually absorb. Second, treat curl's deliberate publication cadence as an operational pattern, not a delay tactic. Third, measure fixes and verified outcomes, not raw findings; a queue that only grows is a warning light, not an asset. Forward-deployed engineering earned its name by staying until the work landed. The 2027 version has to stay until the triage lands too.