Live Intelligence Feed
Continuous Threat Intelligence
Real-time structured feed of jailbreak discoveries, governance updates, detection improvements, and provider advisories. Enterprise intelligence — not social media.
Multilingual role-switch bypass via Bahasa code-switching
Adversary alternates between English and Bahasa Indonesia mid-sentence to defeat English-only moderation. Full system prompt extraction confirmed across 3 enterprise deployments.
Detection model v4.2.1 deployed — improved Aadhaar masking
StreetMP detection model updated to improve Indian Aadhaar number recognition across transliterated and partial-format variants. False positive rate reduced by 34%.
Base64 double-encode circumvents token-level PII scan
Researchers confirm Base64 double-encoding of PII payloads bypasses token-level scan buffers. Model decodes internally and includes raw PII in structured response output.
APAC AI Advisory: Increased prompt injection attempts in healthcare sector
StreetMP telemetry shows 180% increase in prompt injection attempts targeting healthcare AI deployments across India, Singapore, and Malaysia over the past 30 days.
Policy Pack v2.3: Unicode evasion defense released
New governance policy pack released covering 14 Unicode zero-width character variants. Compatible with all StreetMP runtime deployments. See enterprise documentation for integration.
Unicode zero-width join sequence bypasses content filter
Inserting U+200D (ZWNJ) between sensitive tokens prevents string-matching content filters from flagging adversarial payloads in Microsoft Copilot enterprise deployments.
New Research: AI Runtime Governance for APAC Financial Services
Published: 40-page compliance implementation guide for banking institutions deploying LLMs under MAS TRM 2021, India PDPB 2023, and APPI 2022 requirements.
System prompt extraction via few-shot induction confirmed
Carefully crafted few-shot examples induce Gemini 1.5 to pattern-match and reveal confidential system prompt fragments. No mitigation currently available from provider.
Malaysian IC (NRIC) detection improved — transliteration coverage
New entity recognition model trained on 4,200 contributed examples from the research community. Detects all 8 known transliteration evasion variants with 94% accuracy.
Multi-step hypothetical framing erodes safety constraints over 5+ turns
Adversary chains hypothetical-framing prompts across extended conversation context to progressively bypass safety constraints. SEC score 7.8. No single-turn detection possible.
Subscribe to Intelligence Alerts
Receive critical threat alerts and governance updates directly. Enterprise delivery only.