Problems
Read the three news summaries (A–C). Then answer all questions A–D. Click choices or type your answers and press Check. Your score appears without revealing the key. Full solutions are provided later.
Passage A — Prompt-Injection Exploit Targets AI Coders (≈130 words)
Security researchers described a new technique nicknamed the “CopyPasta License Attack.” It hides malicious instructions inside markdown comments placed in files that developers rarely question, such as LICENSE.txt or README.md. When an AI coding assistant reads these files, it may treat the text as authoritative and replicate hidden directions across new or edited files, allowing the rogue content to propagate through a codebase. The disclosure drew attention because a major crypto exchange said its engineers widely use a popular AI-first editor. Experts recommend stronger safeguards: scan repositories for hidden comments, restrict what assistants can change, and require human review before committing AI-written changes. The broader lesson is to treat any untrusted text that enters an AI’s context as potentially hostile, especially in semi‑autonomous “vibe coding” workflows.
Passage B — Study: Faster Shipping, But Many More Risks (≈120 words)
An application‑security firm analyzed tens of thousands of repositories at very large enterprises to understand the impact of AI assistants on software quality. It concluded that developers using assistants produced much more code but also triggered a dramatic rise in “security findings.” Compared with late 2024, the monthly number of issues in mid‑2025 was about ten times higher in the studied set. While syntax mistakes and some logic bugs dropped, deeper problems multiplied: privilege‑escalation paths increased, design flaws appeared more often, and oversized pull requests complicated reviews. The message for leaders was blunt: if you scale AI‑generated code, you must scale application security practices in parallel—or you risk accelerating both delivery and danger.
Passage C — Adoption Surges in India, With Oversight (≈120 words)
A recent multi‑country survey reported strong uptake of AI‑assisted coding among Indian technology leaders. Nearly all respondents said their organizations use such tools, and most expect usage to grow further over the next year. Reported benefits include faster prototyping and productivity gains. Yet the same leaders emphasized governance: almost all said AI‑generated code undergoes peer review before release, and many warned about over‑reliance that could weaken skills, particularly for junior staff. Executives highlighted the need for training and clear policies so engineers can leverage assistants without outsourcing judgment. In short, companies are optimistic about AI‑enabled development but are pairing adoption with controls to keep maintainability and security at acceptable levels.
A) Reading Comprehension — 5 Questions
What is the central risk described in Passage A?
According to Passage B, how did the number of security findings change from late 2024 to mid‑2025 in the observed data?
Which practice best reflects how Indian tech leaders balance AI adoption with risk?
In Passage A, the word “propagate” most nearly means:
“Apiiro’s analysis found that syntax errors increased by 76% when developers used AI assistants.”
B) Vocabulary & Collocation — 5 Questions
over‑reliance most nearly means:
To safeguard code means to:
governance refers to:
Complete the collocation: “peer .”
Security teams examined a new “attack .”
C) Grammar Cloze — 2 Questions
“AI assistants, are embedded in many IDEs, can boost speed but also multiply risks.”
“If teams all untrusted data as potentially malicious, they would reduce prompt‑injection risk.”
D) Speaking Prompt — 30 seconds
Prompt: In light of A–C, should companies move quickly into “vibe coding” with AI agents, or slow down for security and skills? State your position and give one pro and one con.
Speak for ~30 seconds. You can record yourself on your device.
Answers & Explanations
A) Reading
- Q1: B — Hidden prompts in common files can make assistants spread malicious code. Passage A explains markdown comments in LICENSE/README can replicate across files.
- Q2: C — Roughly 10× rise. The study saw a tenfold increase in monthly security findings by mid‑2025 versus late 2024.
- Q3: A — Peer review before release. Passage C highlights widespread oversight and review policies.
- Q4: C — “Propagate” ≈ “spread.” It describes how hidden instructions move through a codebase.
- Q5: False — Syntax errors decreased, while deeper risks increased. Passage B contrasts fewer surface errors with more serious issues.
B) Vocabulary & Collocation
- V1: B — over‑reliance: depending on something too much. Example: “Over‑reliance on code suggestions can weaken debugging skills.”
- V2: A — safeguard: protect from harm. Example: “They added pre‑commit checks to safeguard production branches.”
- V3: D — governance: rules and oversight for how something is run. Example: “Clear AI governance sets who approves model updates.”
- V4: peer review — natural use: “All AI changes go through peer review before merging.”
- V5: attack vector — natural use: “Hidden license text became a new attack vector.”
C) Grammar
- G1: which — Relative pronoun for non‑restrictive clause; commas show extra information.
- G2: treated — 2nd conditional (“If + past, would + base”) to express a hypothetical policy and its result.
D) Speaking — Sample Outline (not a full script)
Hook: “AI can write code faster than ever—so can attackers.” → Point: Adopt vibe coding where it adds value. → Evidence: Adoption is rising and speeds improve, but risks (10× findings; prompt‑injection) are real. → Counter: Slowing down entirely misses productivity gains and learning effects. → Close: Move fast with guardrails: peer review, restricted permissions, scanning, and training.
Sources (past 48 hours, JST)
- “Coinbase’s Go-To AI Coding Tool Found Vulnerable to ‘CopyPasta’ Exploit” — CoinDesk — https://www.coindesk.com/tech/2025/09/06/coinbase-s-go-to-ai-coding-tool-found-vulnerable-to-copypasta-exploit — 2025-09-06 (JST)
- “AI code assistants make developers more efficient at creating security problems” — The Register — https://www.theregister.com/2025/09/05/ai_code_assistants_security_problems/ — 2025-09-05 (JST)
- “AI-assisted coding rises among Indian tech leaders: Report” — The Economic Times (ETtech) — https://economictimes.indiatimes.com/tech/artificial-intelligence/ai-assisted-coding-rises-among-indian-tech-leaders-report/articleshow/123734347.cms — 2025-09-06 (JST)
All reporting above is paraphrased to respect copyright. Dates converted to Japan Standard Time (UTC+9) where possible.