A concise, fact-based update for security and risk professionals. Core security news first, followed by broader risk and policy developments.

🔐 Core Security Intelligence

1) GitHub Copilot Chat flaw enabled remote prompt injection and data leakage

What’s new:
Legit Security disclosed a Copilot Chat vulnerability that combined a CSP bypass with remote prompt injection. The exploit could leak secrets from private repositories (for example, AWS keys) and allow attackers to steer Copilot’s responses. GitHub has addressed the issue.
Source: SecurityWeek

Why it matters:
Developer-adjacent AI assistants sit on the fault line between code, tokens, and CI/CD secrets. A single prompt-injection foothold can escalate into repository data exposure and poisoned guidance across teams.

Defenses:

• Lock down assistant visibility. Scope tokens and repository access for AI extensions to the minimum required, and separate secrets from source wherever possible. Rotate exposed credentials immediately after any suspected leakage.
• Harden developer environments. Enforce Content Security Policy and trusted domains in IDE webviews. Block remote content in documentation or markdown files that could carry hidden injection payloads.
• Instrument assistant activity. Log assistant prompts and responses for critical projects, and alert on anomalous repository reads or mass file access events.

Expert Insight:
This incident shows how a seemingly benign IDE feature can become an exfiltration channel. Treat AI coding assistants as privileged applications subject to the same controls as build pipelines. Expect more cross-channel attacks that use documentation and context windows as covert entry points.

2) Discord breach exposed government ID images and user PII

What’s new:
Discord confirmed a breach affecting roughly 70,000 users. Attackers claim they stole more than two million government-ID images submitted for age verification, though the investigation continues.
Source: SecurityWeek

Why it matters:
Age-verification systems store highly sensitive documents and metadata. A compromise here poses long-term identity and fraud risks that outlast normal credential resets.

Defenses:

• Minimize retention and segment storage. Keep verification images in isolated, short-lived storage with strict access control and encryption. Require deletion proofs from third-party processors.
• Support affected users. Enforce MFA, issue password resets, and watch for synthetic-identity fraud attempts or SIM-swap follow-ups.
• Rehearse incident response. Run tabletop drills for verification-data exposure to ensure breach-notification, regulator contact, and evidence-preservation plans are tested.

Expert Insight:
Verification stacks are compliance features that often lack full-scale security architecture. Make them ephemeral and auditable by default—treat ID imagery like payment data, not generic uploads.

3) DeepMind details agent that finds and fixes vulnerabilities autonomously

What’s new:
Google DeepMind unveiled an AI agent that autonomously discovers and patches vulnerabilities. The system, internally codenamed “Big Sleep,” reportedly found a critical SQLite flaw and helped block exploitation attempts in the wild.
Source: SecurityWeek

Why it matters:
Automated discovery and remediation drastically shrink attacker dwell time but can also inject instability or new bugs faster than humans can review.

Defenses:

• Stage automated fixes. Require canary releases and automated tests before broad deployment, and maintain rollback capabilities for agent-authored patches.
• Restrict agent privileges. Limit write access to non-production repositories and enforce human review before merges into sensitive branches.
• Track post-patch metrics. Watch for regressions, dependency shifts, or new vulnerabilities introduced by automatic remediation.

Expert Insight:
Autonomous vulnerability management is powerful yet risky. Treat these systems like tireless junior engineers: verify every change, test thoroughly, and keep humans in the approval loop. Automation without oversight converts speed into volatility.

4) Russia’s AI-driven cyber operations against Ukraine surge in 2025

What’s new:
Ukraine’s SSSCIP reported more than 3,000 AI-assisted attacks in the first half of 2025, citing AI use for phishing, malware creation, and operational coordination by Russian threat groups.
Source: The Hacker News

Why it matters:
Nation-state adoption of generative AI has lowered costs and broadened operational reach. Campaigns now tailor language, timing, and payloads dynamically, defeating static defenses.

Defenses:

• Bolster identity and messaging security. Deploy phishing-resistant MFA and DKIM/DMARC enforcement, and train staff to recognize AI-crafted content cues.
• Accelerate threat-intel sharing. Exchange prompt fingerprints, payload samples, and model indicators with trusted partners to shorten response cycles.
• Prepare crisis-communication plans. Include verification channels for deepfake or misinformation incidents that may coincide with technical attacks.

Expert Insight:
Hybrid conflict increasingly merges code and content. Even non-government enterprises can become collateral targets in narrative or disruption campaigns. Continuous monitoring and verified communication channels are key defenses.

5) LOKI ML system uncovers 52 000 scam sites via risky search queries

What’s new:
Boston University researchers introduced LOKI, a machine-learning system that correlates risky search queries with scam domains. From 1 663 seed sites, LOKI discovered 52 493 new fraudulent domains across ten categories.
Source: Help Net Security

Why it matters:
The model shows how AI can proactively map and disrupt fraud ecosystems by predicting where scams will appear rather than chasing takedowns afterward.

Defenses:

• Integrate discoveries into filtering tools. Feed new domain indicators into DNS and web-proxy blocklists, focusing on categories relevant to your workforce.
• Correlate with internal telemetry. Combine LOKI-style external data with clickstream and Safe Browsing logs to detect exposure earlier.
• Coordinate on takedowns. Share indicators with registrars and ISPs to cut down reuse and fast-flux re-registration cycles.

Expert Insight:
Attackers exploit curiosity and urgency at the search stage. Mapping intent-driven risk gives defenders an early intercept point. Expect further academic-industry collaboration on predictive scam detection.

🌐 Extended Reading / Broader AI Risk & Governance

Taiwan defense report warns of “hybrid” online warfare tactics

What’s new:
Taiwan’s defense ministry warned that China is refining capabilities for surprise attack and hybrid online operations blending disinformation with cyber intrusion.
Source: Reuters

Why it matters:
Hybrid warfare fuses physical, cyber, and informational domains. Multinational organizations should evaluate regional exposure and supplier resilience under escalating geopolitical tension.

⚠️ Updates / Follow-ups

No new follow-ups were confirmed in the past 24 hours.

Summary Table

Categories: Cybersecurity News

Tags: Agentic AI, AI Security, Data Leakage, Disinformation, Fraud Detection, Prompt Injection