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.

Filter
JAILBREAKCRITICAL GPT-4o
14d ago

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.

#jailbreak#multilingual#gpt-4o#role-switch
by NUS_SecLabView details
DETECTION UPDATE
14d ago

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%.

#detection-update#aadhaar#india#pii
DLP BYPASSHIGH Claude 3.5
14d ago

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.

#dlp-bypass#base64#claude#pii
by r_vasquezView details
ADVISORYHIGH All Providers
14d ago

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.

#advisory#healthcare#apac#injection
MITIGATION
15d ago

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.

#policy-pack#unicode#mitigation
INJECTIONHIGH Copilot
15d ago

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.

#injection#unicode#copilot#dlp
by iitb_aiView details
PAPER
15d ago

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.

#paper#governance#apac#financial-services
by kyoto_secView details
GOVERNANCEMEDIUM Gemini 1.5
16d ago

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.

#governance#few-shot#gemini#extraction
by NUS_SecLabView details
DETECTION UPDATE
16d ago

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.

#detection-update#malaysia#nric#pii
JAILBREAKHIGH Llama 3
17d ago

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.

#jailbreak#semantic#llama#multi-step
by anon_r847View details

Subscribe to Intelligence Alerts

Receive critical threat alerts and governance updates directly. Enterprise delivery only.