Activity

Includes 47 new Hinglish jailbreak variants and 12 Devanagari-encoding attack patterns.

All 7 requirements met; 1 blocked on open critical findings — escalated to Catherine O'Brien.

Refusal-bypass via Devanagari-script encoding — reproducible in 3 of 12 attempts.

Approved with condition: monthly post-deployment monitoring eval + quarterly red-team.

8,420 labelled test cases covering 24 high-risk scenarios across BFSI use-cases.

Approved for use under EU AI Act high-risk category. Conditions: contractual SOC 2 attestation refresh annually.

Auto-curated from claims-copilot-v3 production traffic — gold-standard candidates for regression suite.

Sourced from public corpora + 12 internally-discovered patterns. Coverage: code-mixing, transliteration, cultural framing.

Targeting 12,500 probes across prompt-injection, RAG-poisoning, tool-abuse categories. ETA 4 hours.

v7 swaps embedding to voyage-3 and adds Hinglish refusal templates. Pre-cert evaluation queued.

Eval gate failed: faithfulness regressed from 96% to 87%. Returned to engineering for root-cause analysis.

All findings traced to a single root cause: indirect prompt injection via attacker-authored RAG document.

Passed at 94.2% (+1.8pp vs v17). 3,200 test cases over judge-LLM gpt-4o-mini. Cost $312.

“Recommend immediate input-sanitization on RAG sources before re-cert. CISO sign-off blocked until resolved.”

Pre-quarter risk re-assessment under SR 11-7. Scope: 4 underlying models, 2 RAG pipelines.

Renewed for 90 days with elevated monitoring frequency.

0 critical, 0 high, 2 medium findings. All 2 mediums queued for triage.

Auto-detected new versions deployed in Production. Pending Trust Lab certification gate.

Hindi 92.1% · Hinglish 88.4% · Tamil 84.2%. Sarvam-1 outperforms Llama 3.1 70B by 4.6pp on Hindi.

All fraud-investigation agents restricted from outbound HTTP except whitelisted internal tools. Effective immediately.

Tool-use coverage 91%; 3 scenarios failed (boundary cases involving multi-step evidence chains).