Anthropic Reveals Identity as AI Bug‑Detection Expert, Yet Struggles to Fix Errors

Anthropic’s claim that Claude Code Security can “find and fix flaws they might otherwise have missed” rests on a research preview that, according to the company, helped its red‑team uncover “over 500 vulnerabilities in production open‑source codebases” using Claude Opus 4.6 — a figure highlighted in Anthropic’s own rollout announcement. Yet the practical impact of those findings appears minimal. Guy Azari, a stealth‑startup founder and former security researcher at Microsoft and Palo Alto Networks, told The Register that of the 500 reported bugs, “only two to three vulnerabilities were fixed,” and noted the absence of any CVE assignments, suggesting the remediation pipeline never materialised.

The discrepancy between detection and remediation is not new in the industry, but Anthropic’s publicity amplifies the gap at a time when the Pentagon is pressing the firm for broader access. Defense Secretary Pete Hegseth met with Anthropic co‑founder Dario Amodei on February 24, reportedly demanding “unrestricted Claude access by Friday” or threatening to invoke the Defense Production Act, which would conscript Anthropic into a supply‑chain role for classified defense systems — the same Claude model already deployed via Palantir in operations such as the capture of Venezuela’s Maduro, per the zecheng report. Anthropic’s usage policy explicitly bans mass surveillance and autonomous weapons, placing the company at a crossroads between its “responsible AI” brand and a potential government‑mandated deployment.

Compounding the pressure, Anthropic unveiled a revised Responsible Scaling Policy (RSP 3.0) the same day. The new policy drops the previous hard line that barred training more powerful models without confirmed safety measures, replacing it with language that development will only be “delayed” if leadership believes Anthropic “leads the AI race AND catastrophic risk” — a shift reported by VentureBeat and TechCrunch. This relaxation aligns with the firm’s recent market‑cap surge; enterprise plugins tied to Claude reportedly added $830 billion in valuation, according to the zecheng post, underscoring the financial incentive to accelerate model capabilities despite lingering safety concerns.

Industry observers note that the core challenge remains the translation of AI‑generated vulnerability reports into actionable patches. Azari, drawing on his tenure at the Microsoft Security Response Center, warned that AI “multiplied” the volume of reports by “100x or 200x” and introduced “a lot of noise because AI assumes that these are vulnerabilities,” echoing The Register’s assessment that discovery is becoming cheaper while validation and patching lag behind. Without a robust workflow to triage, prioritize, and assign CVEs, the promise of AI‑driven bug hunting risks inflating false positives and overwhelming security teams, a risk that could erode confidence among enterprise customers.

The confluence of a high‑stakes Pentagon ultimatum, a softened safety policy, and a market‑driven push for enterprise plugins creates a precarious strategic landscape for Anthropic. If the firm cannot demonstrably close the loop from detection to remediation, its positioning as the “AI bug‑detection specialist” may prove more rhetorical than operational, potentially inviting regulatory scrutiny and undermining the very “responsible AI” narrative it seeks to protect.